NAME
stap - systemtap script translator/driver
SYNOPSIS
stap [ OPTIONS ] FILENAME [ ARGUMENTS ]
stap [ OPTIONS ] - [ ARGUMENTS ]
stap [ OPTIONS ] -e SCRIPT [ ARGUMENTS ]
stap [ OPTIONS ] -l PROBE [ ARGUMENTS ]
stap [ OPTIONS ] -L PROBE [ ARGUMENTS ]
DESCRIPTION
The stap program is the front-end to the Systemtap tool. It accepts
probing instructions (written in a simple scripting language),
translates those instructions into C code, compiles this C code, and
loads the resulting kernel module into a running Linux kernel to
perform the requested system trace/probe functions. You can supply the
script in a named file, from standard input, or from the command line.
The program runs until it is interrupted by the user, or if the script
voluntarily invokes the exit() function, or by sufficient number of
soft errors.
The language, which is described in a later section, is strictly typed,
declaration free, procedural, and inspired by awk. It allows source
code points or events in the kernel to be associated with handlers,
which are subroutines that are executed synchronously. It is somewhat
similar conceptually to "breakpoint command lists" in the gdb debugger.
This manual corresponds to version 1.2.
OPTIONS
The systemtap translator supports the following options. Any other
option prints a list of supported options.
-h Show help message.
-V Show version message.
-p NUM Stop after pass NUM. The passes are numbered 1-5: parse,
elaborate, translate, compile, run. See the PROCESSING section
for details.
-v Increase verbosity for all passes. Produce a larger volume of
informative (?) output each time option repeated.
--vp ABCDE
Increase verbosity on a per-pass basis. For example, "--vp 002"
adds 2 units of verbosity to pass 3 only. The combination
"-v --vp 00004" adds 1 unit of verbosity for all passes, and 4
more for pass 5.
-k Keep the temporary directory after all processing. This may be
useful in order to examine the generated C code, or to reuse the
compiled kernel object.
-g Guru mode. Enable parsing of unsafe expert-level constructs
like embedded C.
-P Prologue-searching mode. Activate heuristics to work around
incorrect debugging information for $target variables.
-u Unoptimized mode. Disable unused code elision during
elaboration.
-w Suppressed warnings mode. Disables all warning messages.
-b Use bulk mode (percpu files) for kernel-to-user data transfer.
-t Collect timing information on the number of times probe executes
and average amount of time spent in each probe.
-sNUM Use NUM megabyte buffers for kernel-to-user data transfer. On a
multiprocessor in bulk mode, this is a per-processor amount.
-I DIR Add the given directory to the tapset search directory. See the
description of pass 2 for details.
-D NAME=VALUE
Add the given C preprocessor directive to the module Makefile.
These can be used to override limit parameters described below.
-B NAME=VALUE
Add the given make directive to the kernel module build’s make
invocation. These can be used to add or override kconfig
options.
-R DIR Look for the systemtap runtime sources in the given directory.
-r /DIR
Build for kernel in given build tree. Can also be set with the
SYSTEMTAP_RELEASE environment variable.
-r RELEASE
Build for kernel in build tree /lib/modules/RELEASE/build. Can
also be set with the SYSTEMTAP_RELEASE environment variable.
-m MODULE
Use the given name for the generated kernel object module,
instead of a unique randomized name. The generated kernel
object module is copied to the current directory.
-d MODULE
Add symbol/unwind information for the given module into the
kernel object module. This may enable symbolic tracebacks from
those modules/programs, even if they do not have an explicit
probe placed into them.
-o FILE
Send standard output to named file. In bulk mode, percpu files
will start with FILE_ (FILE_cpu with -F) followed by the cpu
number. This supports strftime(3) formats for FILE.
-c CMD Start the probes, run CMD, and exit when CMD finishes.
-x PID Sets target() to PID. This allows scripts to be written that
filter on a specific process.
-l PROBE
Instead of running a probe script, just list all available probe
points matching the given pattern. The pattern may include
wildcards and aliases.
-L PROBE
Similar to "-l", but list probe points and script-level local
variables.
-F Without -o option, load module and start probes, then detach
from the module leaving the probes running. With -o option, run
staprun in background as a daemon and show its pid.
-S size[,N]
Sets the maximum size of output file and the maximum number of
output files. If the size of output file will exceed size ,
systemtap switches output file to the next file. And if the
number of output files exceed N , systemtap removes the oldest
output file. You can omit the second argument.
--skip-badvars
Ignore out of context variables and substitute with literal 0.
ARGUMENTS
Any additional arguments on the command line are passed to the script
parser for substitution. See below.
SCRIPT LANGUAGE
The systemtap script language resembles awk. There are two main
outermost constructs: probes and functions. Within these, statements
and expressions use C-like operator syntax and precedence.
GENERAL SYNTAX
Whitespace is ignored. Three forms of comments are supported:
# ... shell style, to the end of line, except for $# and @#
// ... C++ style, to the end of line
/* ... C style ... */
Literals are either strings enclosed in double-quotes (passing through
the usual C escape codes with backslashes), or integers (in decimal,
hexadecimal, or octal, using the same notation as in C). All strings
are limited in length to some reasonable value (a few hundred bytes).
Integers are 64-bit signed quantities, although the parser also accepts
(and wraps around) values above positive 2**63.
In addition, script arguments given at the end of the command line may
be inserted. Use $1 ... $<NN> for insertion unquoted, @1 ... @<NN> for
insertion as a string literal. The number of arguments may be accessed
through $# (as an unquoted number) or through @# (as a quoted number).
These may be used at any place a token may begin, including within the
preprocessing stage. Reference to an argument number beyond what was
actually given is an error.
PREPROCESSING
A simple conditional preprocessing stage is run as a part of parsing.
The general form is similar to the cond ? exp1 : exp2 ternary operator:
%( CONDITION %? TRUE-TOKENS %)
%( CONDITION %? TRUE-TOKENS %: FALSE-TOKENS %)
The CONDITION is either an expression whose format is determined by its
first keyword, or a string literals comparison or a numeric literals
comparison. It can be also composed of many alternatives and
conjunctions of CONDITIONs (meant as in previous sentence) using || and
&& respectively. However, parentheses are not supported yet, so
remembering that conjunction takes precedence over alternative is
important.
If the first part is the identifier kernel_vr or kernel_v to refer to
the kernel version number, with ("2.6.13-1.322FC3smp") or without
("2.6.13") the release code suffix, then the second part is one of the
six standard numeric comparison operators <, <=, ==, !=, >, and >=, and
the third part is a string literal that contains an RPM-style version-
release value. The condition is deemed satisfied if the version of the
target kernel (as optionally overridden by the -r option) compares to
the given version string. The comparison is performed by the glibc
function strverscmp. As a special case, if the operator is for simple
equality (==), or inequality (!=), and the third part contains any
wildcard characters (* or ? or [), then the expression is treated as a
wildcard (mis)match as evaluated by fnmatch.
If, on the other hand, the first part is the identifier arch to refer
to the processor architecture (as named by the kernel build system
ARCH/SUBARCH), then the second part is one of the two string comparison
operators == or !=, and the third part is a string literal for matching
it. This comparison is a wildcard (mis)match.
Similarly, if the first part is an identifier like CONFIG_something to
refer to a kernel configuration option, then the second part is == or
!=, and the third part is a string literal for matching the value
(commonly "y" or "m"). Nonexistent or unset kernel configuration
options are represented by the empty string. This comparison is also a
wildcard (mis)match.
Otherwise, the CONDITION is expected to be a comparison between two
string literals or two numeric literals. In this case, the arguments
are the only variables usable.
The TRUE-TOKENS and FALSE-TOKENS are zero or more general parser tokens
(possibly including nested preprocessor conditionals), and are passed
into the input stream if the condition is true or false. For example,
the following code induces a parse error unless the target kernel
version is newer than 2.6.5:
%( kernel_v <= "2.6.5" %? **ERROR** %) # invalid token sequence
The following code might adapt to hypothetical kernel version drift:
probe kernel.function (
%( kernel_v <= "2.6.12" %? "__mm_do_fault" %:
%( kernel_vr == "2.6.13*smp" %? "do_page_fault" %:
UNSUPPORTED %) %)
) { /* ... */ }
%( arch == "ia64" %?
probe syscall.vliw = kernel.function("vliw_widget") {}
%)
VARIABLES
Identifiers for variables and functions are an alphanumeric sequence,
and may include "_" and "$" characters. They may not start with a
plain digit, as in C. Each variable is by default local to the probe
or function statement block within which it is mentioned, and therefore
its scope and lifetime is limited to a particular probe or function
invocation.
Scalar variables are implicitly typed as either string or integer.
Associative arrays also have a string or integer value, and a tuple of
strings and/or integers serving as a key. Here are a few basic
expressions.
var1 = 5
var2 = "bar"
array1 [pid()] = "name" # single numeric key
array2 ["foo",4,i++] += 5 # vector of string/num/num keys
if (["hello",5,4] in array2) println ("yes") # membership test
The translator performs type inference on all identifiers, including
array indexes and function parameters. Inconsistent type-related use
of identifiers signals an error.
Variables may be declared global, so that they are shared amongst all
probes and live as long as the entire systemtap session. There is one
namespace for all global variables, regardless of which script file
they are found within. A global declaration may be written at the
outermost level anywhere, not within a block of code. Global variables
which are written but never read will be displayed automatically at
session shutdown. The translator will infer for each its value type,
and if it is used as an array, its key types. Optionally, scalar
globals may be initialized with a string or number literal. The
following declaration marks variables as global.
global var1, var2, var3=4
Global variables can also be set as module options. To do this, the
module must first be compiled using stap -p4. Global variables can then
be set on the command line when calling staprun on the module generated
by stap -p4. See staprun(8) for more information.
Arrays are limited in size by the MAXMAPENTRIES variable -- see the
SAFETY AND SECURITY section for details. Optionally, global arrays may
be declared with a maximum size in brackets, overriding MAXMAPENTRIES
for that array only. Note that this doesn’t indicate the type of keys
for the array, just the size.
global tiny_array[10], normal_array, big_array[50000]
STATEMENTS
Statements enable procedural control flow. They may occur within
functions and probe handlers. The total number of statements executed
in response to any single probe event is limited to some number defined
by a macro in the translated C code, and is in the neighbourhood of
1000.
EXP Execute the string- or integer-valued expression and throw away
the value.
{ STMT1 STMT2 ... }
Execute each statement in sequence in this block. Note that
separators or terminators are generally not necessary between
statements.
; Null statement, do nothing. It is useful as an optional
separator between statements to improve syntax-error detection
and to handle certain grammar ambiguities.
if (EXP) STMT1 [ else STMT2 ]
Compare integer-valued EXP to zero. Execute the first (non-
zero) or second STMT (zero).
while (EXP) STMT
While integer-valued EXP evaluates to non-zero, execute STMT.
for (EXP1; EXP2; EXP3) STMT
Execute EXP1 as initialization. While EXP2 is non-zero, execute
STMT, then the iteration expression EXP3.
foreach (VAR in ARRAY [ limit EXP ]) STMT
Loop over each element of the named global array, assigning
current key to VAR. The array may not be modified within the
statement. By adding a single + or - operator after the VAR or
the ARRAY identifier, the iteration will proceed in a sorted
order, by ascending or descending index or value. Using the
optional limit keyword limits the number of loop iterations to
EXP times. EXP is evaluated once at the beginning of the loop.
foreach ([VAR1, VAR2, ...] in ARRAY [ limit EXP ]) STMT
Same as above, used when the array is indexed with a tuple of
keys. A sorting suffix may be used on at most one VAR or ARRAY
identifier.
break, continue
Exit or iterate the innermost nesting loop (while or for or
foreach) statement.
return EXP
Return EXP value from enclosing function. If the function’s
value is not taken anywhere, then a return statement is not
needed, and the function will have a special "unknown" type with
no return value.
next Return now from enclosing probe handler. This is especially
useful in probe aliases that apply event filtering predicates.
try { STMT1 } catch { STMT2 }
Run the statements in the first block. Upon any run-time
errors, abort STMT1 and start executing STMT2. Any errors in
STMT2 will propagate to outer try/catch blocks, if any.
try { STMT1 } catch(VAR) { STMT2 }
Same as above, plus assign the error message to the string
scalar variable VAR.
delete ARRAY[INDEX1, INDEX2, ...]
Remove from ARRAY the element specified by the index tuple. The
value will no longer be available, and subsequent iterations
will not report the element. It is not an error to delete an
element that does not exist.
delete ARRAY
Remove all elements from ARRAY.
delete SCALAR
Removes the value of SCALAR. Integers and strings are cleared
to 0 and "" respectively, while statistics are reset to the
initial empty state.
EXPRESSIONS
Systemtap supports a number of operators that have the same general
syntax, semantics, and precedence as in C and awk. Arithmetic is
performed as per typical C rules for signed integers. Division by zero
or overflow is detected and results in an error.
binary numeric operators
* / % + - >> << & ^ | && ||
binary string operators
. (string concatenation)
numeric assignment operators
= *= /= %= += -= >>= <<= &= ^= |=
string assignment operators
= .=
unary numeric operators
+ - ! ~ ++ --
binary numeric or string comparison operators
< > <= >= == !=
ternary operator
cond ? exp1 : exp2
grouping operator
( exp )
function call
fn ([ arg1, arg2, ... ])
array membership check
exp in array
[exp1, exp2, ...] in array
PROBES
The main construct in the scripting language identifies probes. Probes
associate abstract events with a statement block ("probe handler") that
is to be executed when any of those events occur. The general syntax
is as follows:
probe PROBEPOINT [, PROBEPOINT] { [STMT ...] }
Events are specified in a special syntax called "probe points". There
are several varieties of probe points defined by the translator, and
tapset scripts may define further ones using aliases. These are listed
in the stapprobes(3stap) manual pages.
The probe handler is interpreted relative to the context of each event.
For events associated with kernel code, this context may include
variables defined in the source code at that spot. These "target
variables" are presented to the script as variables whose names are
prefixed with "$". They may be accessed only if the kernel’s compiler
preserved them despite optimization. This is the same constraint that
a debugger user faces when working with optimized code. Some other
events have very little context. See the stapprobes(3stap) man pages
to see the kinds of context variables available at each kind of probe
point.
New probe points may be defined using "aliases". Probe point aliases
look similar to probe definitions, but instead of activating a probe at
the given point, it just defines a new probe point name as an alias to
an existing one. There are two types of alias, i.e. the prologue style
and the epilogue style which are identified by "=" and "+="
respectively.
For prologue style alias, the statement block that follows an alias
definition is implicitly added as a prologue to any probe that refers
to the alias. While for the epilogue style alias, the statement block
that follows an alias definition is implicitly added as an epilogue to
any probe that refers to the alias. For example:
probe syscall.read = kernel.function("sys_read") {
fildes = $fd
if (execname == "init") next # skip rest of probe
}
defines a new probe point syscall.read, which expands to
kernel.function("sys_read"), with the given statement as a prologue,
which is useful to predefine some variables for the alias user and/or
to skip probe processing entirely based on some conditions. And
probe syscall.read += kernel.function("sys_read") {
if (tracethis) println ($fd)
}
defines a new probe point with the given statement as an epilogue,
which is useful to take actions based upon variables set or left over
by the the alias user.
An alias is used just like a built-in probe type.
probe syscall.read {
printf("reading fd=%d0, fildes)
if (fildes > 10) tracethis = 1
}
FUNCTIONS
Systemtap scripts may define subroutines to factor out common work.
Functions take any number of scalar (integer or string) arguments, and
must return a single scalar (integer or string). An example function
declaration looks like this:
function thisfn (arg1, arg2) {
return arg1 + arg2
}
Note the general absence of type declarations, which are instead
inferred by the translator. However, if desired, a function definition
may include explicit type declarations for its return value and/or its
arguments. This is especially helpful for embedded-C functions. In
the following example, the type inference engine need only infer type
type of arg2 (a string).
function thatfn:string (arg1:long, arg2) {
return sprint(arg1) . arg2
}
Functions may call others or themselves recursively, up to a fixed
nesting limit. This limit is defined by a macro in the translated C
code and is in the neighbourhood of 10.
PRINTING
There are a set of function names that are specially treated by the
translator. They format values for printing to the standard systemtap
output stream in a more convenient way. The sprint* variants return
the formatted string instead of printing it.
print, sprint
Print one or more values of any type, concatenated directly
together.
println, sprintln
Print values like print and sprint, but also append a newline.
printd, sprintd
Take a string delimiter and two or more values of any type, and
print the values with the delimiter interposed. The delimiter
must be a literal string constant.
printdln, sprintdln
Print values with a delimiter like printd and sprintd, but also
append a newline.
printf, sprintf
Take a formatting string and a number of values of corresponding
types, and print them all. The format must be a literal string
constant.
The printf formatting directives similar to those of C, except that
they are fully type-checked by the translator:
%b Writes a binary blob of the value given, instead of ASCII
text. The width specifier determines the number of bytes
to write; valid specifiers are %b %1b %2b %4b %8b.
Default (%b) is 8 bytes.
%c Character.
%d,%i Signed decimal.
%m Safely reads kernel memory at the given address, outputs
its content. The precision specifier determines the
number of bytes to read. Default is 1 byte.
%M Same as %m, but outputs in hexadecimal. The minimal size
of output is double the precision specifier.
%o Unsigned octal.
%p Unsigned pointer address.
%s String.
%u Unsigned decimal.
%x Unsigned hex value, in all lower-case.
%X Unsigned hex value, in all upper-case.
%% Writes a %.
Examples:
a = "alice", b = "bob", p = 0x1234abcd, i = 123, j = -1, id[a] = 1234, id[b] = 4567
print("hello")
Prints: hello
println(b)
Prints: bob\n
println(a . " is " . sprint(16))
Prints: alice is 16
foreach (name in id) printdln("|", strlen(name), name, id[name])
Prints: 5|alice|1234\n3|bob|4567
printf("%c is %s; %x or %X or %p; %d or %u\n",97,a,p,p,p,j,j)
Prints: a is alice; 1234abcd or 1234ABCD or 0x1234abcd; -1 or 18446744073709551615\n
printf("2 bytes of kernel buffer at address %p: %2m", p, p)
Prints: 2 byte of kernel buffer at address 0x1234abcd: <binary data>
printf("%4b", p)
Prints (these values as binary data): 0x1234abcd
STATISTICS
It is often desirable to collect statistics in a way that avoids the
penalties of repeatedly exclusive locking the global variables those
numbers are being put into. Systemtap provides a solution using a
special operator to accumulate values, and several pseudo-functions to
extract the statistical aggregates.
The aggregation operator is <<<, and resembles an assignment, or a C++
output-streaming operation. The left operand specifies a scalar or
array-index lvalue, which must be declared global. The right operand
is a numeric expression. The meaning is intuitive: add the given
number to the pile of numbers to compute statistics of. (The specific
list of statistics to gather is given separately, by the extraction
functions.)
foo <<< 1
stats[pid()] <<< memsize
The extraction functions are also special. For each appearance of a
distinct extraction function operating on a given identifier, the
translator arranges to compute a set of statistics that satisfy it.
The statistics system is thereby "on-demand". Each execution of an
extraction function causes the aggregation to be computed for that
moment across all processors.
Here is the set of extractor functions. The first argument of each is
the same style of lvalue used on the left hand side of the accumulate
operation. The @count(v), @sum(v), @min(v), @max(v), @avg(v) extractor
functions compute the number/total/minimum/maximum/average of all
accumulated values. The resulting values are all simple integers.
Histograms are also available, but are more complicated because they
have a vector rather than scalar value.
@hist_linear(v,start,stop,interval) represents a linear histogram from
"start" to "stop" by increments of "interval". The interval must be
positive. Similarly, @hist_log(v) represents a base-2 logarithmic
histogram. Printing a histogram with the print family of functions
renders a histogram object as a tabular "ASCII art" bar chart.
probe foo {
x <<< $value
}
probe end {
printf ("avg %d = sum %d / count %d\n",
@avg(x), @sum(x), @count(x))
print (@hist_log(v))
}
TYPECASTING
Once a pointer has been saved into a script integer variable, the
translator loses the type information necessary to access members from
that pointer. Using the @cast() operator tells the translator how to
read a pointer.
@cast(p, "type_name"[, "module"])->member
This will interpret p as a pointer to a struct/union named type_name
and dereference the member value. Further ->subfield expressions may
be appended to dereference more levels. NOTE: the same dereferencing
operator -> is used to refer to both direct containment or pointer
indirection. Systemtap automatically determines which. The optional
module tells the translator where to look for information about that
type. Multiple modules may be specified as a list with : separators.
If the module is not specified, it will default either to the probe
module for dwarf probes, or to "kernel" for functions and all other
probes types.
The translator can create its own module with type information from a
header surrounded by angle brackets, in case normal debuginfo is not
available. For kernel headers, prefix it with "kernel" to use the
appropriate build system. All other headers are build with default GCC
parameters into a user module. Multiple headers may be specified in
sequence to resolve a codependency.
@cast(tv, "timeval", "<sys/time.h>")->tv_sec
@cast(task, "task_struct", "kernel<linux/sched.h>")->tgid
@cast(task, "task_struct",
"kernel<linux/sched.h><linux/fs_struct.h>")->fs->umask
When in guru mode, the translator will also allow scripts to assign new
values to members of typecasted pointers.
Typecasting is also useful in the case of void* members whose type may
be determinable at runtime.
probe foo {
if ($var->type == 1) {
value = @cast($var->data, "type1")->bar
} else {
value = @cast($var->data, "type2")->baz
}
print(value)
}
EMBEDDED C
When in guru mode, the translator accepts embedded code in the script.
Such code is enclosed between %{ and %} markers, and is transcribed
verbatim, without analysis, in some sequence, into the generated C
code. At the outermost level, this may be useful to add #include
instructions, and any auxiliary definitions for use by other embedded
code.
The other place where embedded code is permitted is as a function body.
In this case, the script language body is replaced entirely by a piece
of C code enclosed again between %{ and %} markers. This C code may do
anything reasonable and safe. There are a number of undocumented but
complex safety constraints on atomicity, concurrency, resource
consumption, and run time limits, so this is an advanced technique.
The memory locations set aside for input and output values are made
available to it using a macro THIS. Here are some examples:
function add_one (val) %{
THIS->__retvalue = THIS->val + 1;
%}
function add_one_str (val) %{
strlcpy (THIS->__retvalue, THIS->val, MAXSTRINGLEN);
strlcat (THIS->__retvalue, "one", MAXSTRINGLEN);
%}
The function argument and return value types have to be inferred by the
translator from the call sites in order for this to work. The user
should examine C code generated for ordinary script-language functions
in order to write compatible embedded-C ones.
BUILT-INS
A set of builtin functions and probe point aliases are provided by the
scripts installed under the /usr/share/systemtap/tapset directory.
These are described in the stapfuncs(3stap) and stapprobes(3stap)
manual pages.
PROCESSING
The translator begins pass 1 by parsing the given input script, and all
scripts (files named *.stp) found in a tapset directory. The
directories listed with -I are processed in sequence, each processed in
"guru mode". For each directory, a number of subdirectories are also
searched. These subdirectories are derived from the selected kernel
version (the -R option), in order to allow more kernel-version-specific
scripts to override less specific ones. For example, for a kernel
version 2.6.12-23.FC3 the following patterns would be searched, in
sequence: 2.6.12-23.FC3/*.stp, 2.6.12/*.stp, 2.6/*.stp, and finally
*.stp Stopping the translator after pass 1 causes it to print the parse
trees.
In pass 2, the translator analyzes the input script to resolve symbols
and types. References to variables, functions, and probe aliases that
are unresolved internally are satisfied by searching through the parsed
tapset scripts. If any tapset script is selected because it defines an
unresolved symbol, then the entirety of that script is added to the
translator’s resolution queue. This process iterates until all symbols
are resolved and a subset of tapset scripts is selected.
Next, all probe point descriptions are validated against the wide
variety supported by the translator. Probe points that refer to code
locations ("synchronous probe points") require the appropriate kernel
debugging information to be installed. In the associated probe
handlers, target-side variables (whose names begin with "$") are found
and have their run-time locations decoded.
Next, all probes and functions are analyzed for optimization
opportunities, in order to remove variables, expressions, and functions
that have no useful value and no side-effect. Embedded-C functions are
assumed to have side-effects unless they include the magic string
/* pure */. Since this optimization can hide latent code errors such
as type mismatches or invalid $target variables, it sometimes may be
useful to disable the optimizations with the -u option.
Finally, all variable, function, parameter, array, and index types are
inferred from context (literals and operators). Stopping the
translator after pass 2 causes it to list all the probes, functions,
and variables, along with all inferred types. Any inconsistent or
unresolved types cause an error.
In pass 3, the translator writes C code that represents the actions of
all selected script files, and creates a Makefile to build that into a
kernel object. These files are placed into a temporary directory.
Stopping the translator at this point causes it to print the contents
of the C file.
In pass 4, the translator invokes the Linux kernel build system to
create the actual kernel object file. This involves running make in
the temporary directory, and requires a kernel module build system
(headers, config and Makefiles) to be installed in the usual spot
/lib/modules/VERSION/build. Stopping the translator after pass 4 is
the last chance before running the kernel object. This may be useful
if you want to archive the file.
In pass 5, the translator invokes the systemtap auxiliary program
staprun program for the given kernel object. This program arranges to
load the module then communicates with it, copying trace data from the
kernel into temporary files, until the user sends an interrupt signal.
Any run-time error encountered by the probe handlers, such as running
out of memory, division by zero, exceeding nesting or runtime limits,
results in a soft error indication. Soft errors in excess of MAXERRORS
block of all subsequent probes (except error-handling probes), and
terminate the session. Finally, staprun unloads the module, and cleans
up.
ABNORMAL TERMINATION
One should avoid killing the stap process forcibly, for example with
SIGKILL, because the stapio process (a child process of the stap
process) and the loaded module may be left running on the system. If
this happens, send SIGTERM or SIGINT to any remaining stapio processes,
then use rmmod to unload the systemtap module.
EXAMPLES
See the stapex(3stap) manual page for a collection of samples.
CACHING
The systemtap translator caches the pass 3 output (the generated C
code) and the pass 4 output (the compiled kernel module) if pass 4
completes successfully. This cached output is reused if the same
script is translated again assuming the same conditions exist (same
kernel version, same systemtap version, etc.). Cached files are stored
in the $SYSTEMTAP_DIR/cache directory. The cache can be limited by
having the file cache_mb_limit placed in the cache directory (shown
above) containing only an ASCII integer representing how many MiB the
cache should not exceed. Note that this is a ’soft’ limit in that the
cache will be cleaned after a new entry is added, so the total cache
size may temporarily exceed this limit. In the absence of this file, a
default will be created with the limit set to 64MiB.
SAFETY AND SECURITY
Systemtap is an administrative tool. It exposes kernel internal data
structures and potentially private user information. It acquires
either root privileges
To actually run the kernel objects it builds, a user must be one of the
following:
· the root user;
· a member of the stapdev group; or
· a member of the stapusr group.
Members of the stapusr group can only use modules under the following
conditions:
· The module is located in the /lib/modules/VERSION/systemtap
directory. This directory must be owned by root and not be world
writable.
· The module has been signed by a trusted signer. Trusted signers are
normally systemtap compile servers which sign modules when the
--unprivileged option is specified by the client. See the stap-
server(8) manual page for a for more information.
The kernel modules generated by stap program are run by the staprun
program. The latter is a part of the Systemtap package, dedicated to
module loading and unloading (but only in the white zone), and kernel-
to-user data transfer. Since staprun does not perform any additional
security checks on the kernel objects it is given, it would be unwise
for a system administrator to add untrusted users to the stapdev or
stapusr groups.
The translator asserts certain safety constraints. It aims to ensure
that no handler routine can run for very long, allocate memory, perform
unsafe operations, or in unintentionally interfere with the kernel.
Use of script global variables is suitably locked to protect against
manipulation by concurrent probe handlers. Use of guru mode constructs
such as embedded C can violate these constraints, leading to kernel
crash or data corruption.
The resource use limits are set by macros in the generated C code.
These may be overridden with the -D flag. A selection of these is as
follows:
MAXNESTING
Maximum number of nested function calls. Default determined by
script analysis, with a bonus 10 slots added for recursive
scripts.
MAXSTRINGLEN
Maximum length of strings, default 128.
MAXTRYLOCK
Maximum number of iterations to wait for locks on global
variables before declaring possible deadlock and skipping the
probe, default 1000.
MAXACTION
Maximum number of statements to execute during any single probe
hit (with interrupts disabled), default 1000.
MAXACTION_INTERRUPTIBLE
Maximum number of statements to execute during any single probe
hit which is executed with interrupts enabled (such as begin/end
probes), default (MAXACTION * 10).
MAXMAPENTRIES
Maximum number of rows in any single global array, default 2048.
MAXERRORS
Maximum number of soft errors before an exit is triggered,
default 0, which means that the first error will exit the
script.
MAXSKIPPED
Maximum number of skipped probes before an exit is triggered,
default 100. Running systemtap with -t (timing) mode gives more
details about skipped probes. With the default
-DINTERRUPTIBLE=1 setting, probes skipped due to reentrancy are
not accumulated against this limit.
MINSTACKSPACE
Minimum number of free kernel stack bytes required in order to
run a probe handler, default 1024. This number should be large
enough for the probe handler’s own needs, plus a safety margin.
MAXUPROBES
Maximum number of concurrently armed user-space probes
(uprobes), default somewhat larger than the number of user-space
probe points named in the script. This pool needs to be
potentialy large because individual uprobe objects (about 64
bytes each) are allocated for each process for each matching
script-level probe.
STP_MAXMEMORY
Maximum amount of memory (in kilobytes) that the systemtap
module should use, default unlimited. The memory size includes
the size of the module itself, plus any additional allocations.
This only tracks direct allocations by the systemtap runtime.
This does not track indirect allocations (as done by
kprobes/uprobes/etc. internals).
STP_PROCFS_BUFSIZE
Size of procfs probe read buffers (in bytes). Defaults to
MAXSTRINGLEN. This value can be overridden on a per-procfs file
basis using the procfs read probe .maxsize(MAXSIZE) parameter.
With scripts that contain probes on any interrupt path, it is possible
that those interrupts may occur in the middle of another probe handler.
The probe in the interrupt handler would be skipped in this case to
avoid reentrance. To work around this issue, execute stap with the
option -DINTERRUPTIBLE=0 to mask interrupts throughout the probe
handler. This does add some extra overhead to the probes, but it may
prevent reentrance for common problem cases. However, probes in NMI
handlers and in the callpath of the stap runtime may still be skipped
due to reentrance.
Multiple scripts can write data into a relay buffer concurrently. A
host script provides an interface for accessing its relay buffer to
guest scripts. Then, the output of the guests are merged into the
output of the host. To run a script as a host, execute stap with
-DRELAYHOST[=name] option. The name identifies your host script among
several hosts. While running the host, execute stap with
-DRELAYGUEST[=name] to add a guest script to the host. Note that you
must unload guests before unloading a host. If there are some guests
connected to the host, unloading the host will be failed.
In case something goes wrong with stap or staprun after a probe has
already started running, one may safely kill both user processes, and
remove the active probe kernel module with rmmod. Any pending trace
messages may be lost.
In addition to the methods outlined above, the generated kernel module
also uses overload processing to make sure that probes can’t run for
too long. If more than STP_OVERLOAD_THRESHOLD cycles (default
500000000) have been spent in all the probes on a single cpu during the
last STP_OVERLOAD_INTERVAL cycles (default 1000000000), the probes have
overloaded the system and an exit is triggered.
By default, overload processing is turned on for all modules. If you
would like to disable overload processing, define STP_NO_OVERLOAD.
FILES
~/.systemtap
Systemtap data directory for cached systemtap files, unless
overridden by the SYSTEMTAP_DIR environment variable.
/tmp/stapXXXXXX
Temporary directory for systemtap files, including translated C
code and kernel object.
/usr/share/systemtap/tapset
The automatic tapset search directory, unless overridden by the
SYSTEMTAP_TAPSET environment variable.
/usr/share/systemtap/runtime
The runtime sources, unless overridden by the SYSTEMTAP_RUNTIME
environment variable.
/lib/modules/VERSION/build
The location of kernel module building infrastructure.
/usr/lib/debug/lib/modules/VERSION
The location of kernel debugging information when packaged into
the kernel-debuginfo RPM, unless overridden by the
SYSTEMTAP_DEBUGINFO_PATH environment variable. The default
value for this variable is +:.debug:/usr/lib/debug:build.
Elfutils searches vmlinux in this path and it interprets the
path as a base directory of which various subdirectories will be
searched for finding modules.
/usr/bin/staprun
The auxiliary program supervising module loading, interaction,
and unloading.
SEE ALSO
stapprobes(3stap), stapfuncs(3stap), staprun(8), stapvars(3stap),
stapex(3stap), stap-client(8), stap-server(8), awk(1), gdb(1)
BUGS
Use the Bugzilla link of the project web page or our mailing list.
http://sources.redhat.com/systemtap/,<systemtap@sources.redhat.com>.