Performance measurements are listed right now as a flat list, which is
fine when we measure big blocks. But when we start adding more and
more measurements, some of them could be just part of a bigger
measurement and a flat list gives a wrong impression that they are
executed at the same level instead of nested.
Add trace_performance_enter() and trace_performance_leave() to allow
indent these nested measurements. For now it does not help much
because the only nested thing is (lazy) name hash initialization
(e.g. called in diff-index from "git status"). This will help more
because I'm going to add some more tracing that's actually nested.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is so that we can print traces based on this key outside trace.c.
Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Move just enough code from trace.c into trace.h header so all code
necessary to determine that trace is disabled could be inlined to
calling functions. Then perform the check if the trace key is
enabled sooner in call chain.
Signed-off-by: Gennady Kupava <gkupava@bloomberg.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Trace key normalization is not used, not strictly necessary,
complicates the code and would negatively affect compilation speed if
moved to header.
New trace_default_key key or existing separate marco could be used
instead of passing NULL as a key.
Signed-off-by: Gennady Kupava <gkupava@bloomberg.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When debugging the pack protocol, it is sometimes useful to
store the verbatim pack that we sent or received on the
wire. Looking at the on-disk result is often not helpful for
a few reasons:
1. If the operation is a clone, we destroy the repo on
failure, leaving nothing on disk.
2. If the pack is small, we unpack it immediately, and the
full pack never hits the disk.
3. If we feed the pack to "index-pack --fix-thin", the
resulting pack has the extra delta bases added to it.
We already have a GIT_TRACE_PACKET mechanism for tracing
packets. Let's extend it with GIT_TRACE_PACKFILE to dump the
verbatim packfile.
There are a few other positive fallouts that come from
rearranging this code:
- We currently disable the packet trace after seeing the
PACK header, even though we may get human-readable lines
on other sidebands; now we include them in the trace.
- We currently try to print "PACK ..." in the trace to
indicate that the packfile has started. But because we
disable packet tracing, we never printed this line. We
will now do so.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use trace_performance to measure and print execution time and command line
arguments of the entire main() function. In constrast to the shell's 'time'
utility, which measures total time of the parent process, this logs all
involved git commands recursively. This is particularly useful to debug
performance issues of scripted commands (i.e. which git commands were
called with which parameters, and how long did they execute).
Due to git's deliberate use of exit(), the implementation uses an atexit
routine rather than just adding trace_performance_since() at the end of
main().
Usage example: > GIT_TRACE_PERFORMANCE=~/git-trace.log git stash list
Creates a log file like this:
23:57:38.638765 trace.c:405 performance: 0.000310107 s: git command: 'git' 'rev-parse' '--git-dir'
23:57:38.644387 trace.c:405 performance: 0.000261759 s: git command: 'git' 'rev-parse' '--show-toplevel'
23:57:38.646207 trace.c:405 performance: 0.000304468 s: git command: 'git' 'config' '--get-colorbool' 'color.interactive'
23:57:38.648491 trace.c:405 performance: 0.000241667 s: git command: 'git' 'config' '--get-color' 'color.interactive.help' 'red bold'
23:57:38.650465 trace.c:405 performance: 0.000243063 s: git command: 'git' 'config' '--get-color' '' 'reset'
23:57:38.654850 trace.c:405 performance: 0.025126313 s: git command: 'git' 'stash' 'list'
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add trace_performance and trace_performance_since macros that print a
duration and an optional printf-formatted text to the file specified in
environment variable GIT_TRACE_PERFORMANCE.
These macros, in conjunction with getnanotime(), are intended to simplify
performance measurements from within the application (i.e. profiling via
manual instrumentation, rather than using an external profiling tool).
Unless enabled via GIT_TRACE_PERFORMANCE, these macros have no noticeable
impact on performance, so that test code for well known time killers may
be shipped in release builds. Alternatively, a developer could provide an
additional performance patch (not meant for master) that allows reviewers
to reproduce performance tests more easily, e.g. on other platforms or
using their own repositories.
Usage examples:
Simple use case (measure one code section):
uint64_t start = getnanotime();
/* code section to measure */
trace_performance_since(start, "foobar");
Complex use case (measure repetitive code sections):
uint64_t t = 0;
for (;;) {
/* ignore */
t -= getnanotime();
/* code section to measure */
t += getnanotime();
/* ignore */
}
trace_performance(t, "frotz");
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add a getnanotime() function that returns nanoseconds since 01/01/1970 as
unsigned 64-bit integer (i.e. overflows in july 2554). This is easier to
work with than e.g. struct timeval or struct timespec. Basing the timer on
the epoch allows using the results with other time-related APIs.
To simplify adaption to different platforms, split the implementation into
a common getnanotime() and a platform-specific highres_nanos() function.
The common getnanotime() function handles errors, falling back to
gettimeofday() if highres_nanos() isn't implemented or doesn't work.
getnanotime() is also responsible for normalizing to the epoch. The offset
to the system clock is calculated only once on initialization, i.e.
manually setting the system clock has no impact on the timer (except if
the fallback gettimeofday() is in use). Git processes are typically short
lived, so we don't need to handle clock drift.
The highres_nanos() function returns monotonically increasing nanoseconds
relative to some arbitrary point in time (e.g. system boot), or 0 on
failure. Providing platform-specific implementations should be relatively
easy, e.g. adapting to clock_gettime() as defined by the POSIX realtime
extensions is seven lines of code.
This version includes highres_nanos() implementations for:
* Linux: using clock_gettime(CLOCK_MONOTONIC)
* Windows: using QueryPerformanceCounter()
Todo:
* enable clock_gettime() on more platforms
* add Mac OSX version, e.g. using mach_absolute_time + mach_timebase_info
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is useful to see where trace output came from.
Add 'const char *file, int line' parameters to the printing functions and
rename them to *_fl.
Add trace_printf* and trace_strbuf macros resolving to the *_fl functions
and let the preprocessor fill in __FILE__ and __LINE__.
As the trace_printf* functions take a variable number of arguments, this
requires variadic macros (i.e. '#define foo(...) foo_impl(__VA_ARGS__)'.
Though part of C99, it is unclear whether older compilers support this.
Thus keep the old functions and only enable variadic macros for GNUC and
MSVC 2005+ (_MSC_VER 1400). This has the nice side effect that the old
C-style declarations serve as documentation how the macros are to be used.
Print 'file:line ' as prefix to each trace line. Align the remaining trace
output at column 40 to accommodate 18 char file names + 4 digit line
number (currently there are 30 *.c files of length 18 and just 11 of 19).
Trace output from longer source files (e.g. builtin/receive-pack.c) will
not be aligned.
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
No functional changes, just move stuff around so that the next patch isn't
that ugly...
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
To be able to add a common prefix or suffix to all trace output (e.g.
a timestamp or file:line of the caller), factor out common setup and
cleanup tasks of the trace* functions.
When adding a common prefix, it makes sense that the output of each trace
call starts on a new line. Add '\n' in case the caller forgot.
Note that this explicitly limits trace output to line-by-line, it is no
longer possible to trace-print just part of a line. Until now, this was
just an implicit assumption (trace-printing part of a line worked, but
messed up the trace file if multiple threads or processes were involved).
Thread-safety / inter-process-safety is also the reason why we need to do
the prefixing and suffixing in memory rather than issuing multiple write()
calls. Write_or_whine_pipe() / xwrite() is atomic unless the size exceeds
MAX_IO_SIZE (8MB, see wrapper.c). In case of trace_strbuf, this costs an
additional string copy (which should be irrelevant for performance in light
of actual file IO).
While we're at it, rename trace_strbuf's 'buf' argument, which suggests
that the function is modifying the buffer. Trace_strbuf() currently is the
only trace API that can print arbitrary binary data (without barfing on
'%' or stopping at '\0'), so 'data' seems more appropriate.
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The trace API currently rechecks the environment variable and reopens the
trace file on every API call. This has the ugly side effect that errors
(e.g. file cannot be opened, or the user specified a relative path) are
also reported on every call. Performance can be improved by about factor
three by remembering the environment state and keeping the file open.
Replace the 'const char *key' parameter in the API with a pointer to a
'struct trace_key' that bundles the environment variable name with
additional, trace-internal state. Change the call sites of these APIs to
use a static 'struct trace_key' instead of a string constant.
In trace.c::get_trace_fd(), save and reuse the file descriptor in 'struct
trace_key'.
Add a 'trace_disable()' API, so that packet_trace() can cleanly disable
tracing when it encounters packed data (instead of using unsetenv()).
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The format parameter to trace_printf functions is sometimes abbreviated
'fmt'. Rename to 'format' everywhere (consistent with POSIX' printf
specification).
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Also include direct dependencies (strbuf.h and git-compat-util.h for
__attribute__) so that trace.h can be used independently of cache.h, e.g.
in test programs.
Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Nguyễn Thái Ngọc Duy
Gennady Kupava
Jeff King
Karsten Blees