Make the qbipcs.h module interdependence clear (also shedding light to
some semantic dependencies) as well.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
... also to make it the documentation refer to the implementation limits
properly.
When at it, also document some nits on the implementation side, unify
qbarray.h with project's doxygen conventions a bit + fix a typo there.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
Amongst others, care to spell out meaning of "QB_LOG_CONF_PRIORITY_BUMP"
and "fsync'ed logs to file targets" a bit.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
The object in question has never been published through the header file,
hence it's presumably safe to make it static as it's meant to be.
On the other hand, QB_LOG_INIT_DATA macro from qblog.h has already
started to depend on that symbol so as to locate the library handle
for libqb itself correctly. This is trivially fixed by finally exposing
library versioning info in run-time ("online") as a structure with
members corresponding to compile-time ("offline") counterparts from
qbconfig.h header file, which are admittedly of very limited use
as opposed to the newly introduced dynamic info, plus lower-cased
equivalent of QB_VER_STR. Better than to roll out a futile data object
serving as an artificial anchor for the above purpose, and this was
due for a while, afterall.
In turn, also bump "current" and "age" of fields of the libtool's
"-version-info" versioning system.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
(or rather [read on]: "bare" fix, now that we established means to
analyse the impact of the linker-dependent misbehaviour and to detect
some of its symptoms in preceding two commits, respectively)
Initially with the help of the internal test suite and the failing log
test, it was eventually discovered[1] that these binutils commits going
to the recent 2.29 release affected the treatment of _start_SECNAME
and __stop_SECNAME symbols denoting the boundary start/stop addresses
of a SECNAME orphan section -- specifically in libqb context a custom
section (SECNAME=__verbose) used for link-time ("run-time amortizing")
callsite collection when there's a support in the toolchain[*]:
https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=cbd0eecf261c2447781f8c89b0d955ee66fae7e9https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=b27685f2016c510d03ac9a64f7b04ce8efcf95c4https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=7dba9362c172f1073487536eb137feb2da30b0ff
The first one explicitly states:
> Also __start_SECNAME and __stop_SECNAME symbols are marked as hidden
> by ELF linker so that __start_SECNAME and __stop_SECNAME symbols for
> section SECNAME in different modules are unique.
The problem is that libqb silently depends on the previous status quo
ld.bfd linker behaviour of keeping those symbols externally visible,
which was apparently not granted as it has deliberately changed per
above.
And then for 2.29.1 release of binutils once again, as someone actually
noticed something went overboard with the 2.29 changes:
http://lists.gnu.org/archive/html/bug-binutils/2017-08/msg00195.html
(overview of the original bug discussion, rather than directly
https://sourceware.org/bugzilla/show_bug.cgi?id=21964, which is
a result of a conflct resolution when restoring bugzilla backup)
https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=487b6440dad57440939fab7afdd84a218b612796
At least that change doesn't invalidate all the effort being put into
the original version of the changeset, only the configure script check
had to be refined so as not to miss the "orphan section magic not
working properly out of the box, without band aid" observation
(see the inline comment) -- the workaround arrangement needs
to be applied in that case as well.
* * *
So regarding the solution itself, the core of the fix was sketched at
the original Fedora targeted bug against binutils[2]. In short, we are
using a custom linker script that (re)describes the mentioned custom
orphan output section, or better yet, assuredly pushes that section, and
more importantly, it's own boundary denoting symbols, through into the
resulting executable when it's being linked (as in compile-time step).
This solution alone, while working for the non-libqb (more on that
below) logging participants, is not good enough, as it requires all
libqb targets to start using new incantation (namely "-Wl,foo.t" switch)
in the final link step during compilation, which might be solvable
with a tweak in libqb's pkg-config file under assumption that practice
of using "pkg-config --libs libqb" is rigidly followed. Which is likely
a false expectation, and furthermore only for the regular consumption
model, as it doesn't cover the least bit the developmental one (refer
to previous-but-one "tests" commit message), e.g. applied for internal
examples + tests (but no local sub-checkout tree usage can be excluded).
So further extensions were devised to cover both consumption models:
- a. regular:
courtesy of binutils maintainer[3], we follow an idea to make libqb.so
(i.e. what the targets link against) rather a linker script on its
own, which first include the version-specified (e.g. libqb.so.0) file
into the link, then lists, in situ, the content of the linker script
per above, hence -lqb linking has the same effect as having both
"-lqb -Wl,foo.t" explicitly in the link command prior to this trick
- b. developmental:
to eliminate any kind of race condition arising from the attempt
to post-modify libqb.la libtool archive file generated internally
by libtool, we sort of abuse "inherited_linker_flags" variable
within this file format, as it forms an accumulative value across
the whole transitive dependencies chain (if not impaired per the
note below), fitting exactly our purpose of injecting "-Wl,foo.t"
switch equivalent for those libtool-linking by L{D,IB}ADD'ing
libqb.la; it's then enough to craft a custom libtool archive file
declaring that value, and hook it into such dependency chain through
libqb_la_LIBADD, and with a little bit of further fiddling, it works
as desired (note that double occurrence of "-Wl,foo.t" equivalent
present at some stages of sorting this trick turned out to be,
surprisingly, counter-productive, which should now demistify the
very existence of effectively empty qblog_script_noop.ld file);
NOTE: some forms of libtool distribution (debian + derivatives ones
in particular) undermine natural transitive dependency propagation
with a deliberate cut off (https://bugs.debian.org/702737), so we
need to ensure the "impairment" is not happening by force (corosync
precedent: https://github.com/corosync/corosync/commit/0f1dc5c1)
^ something like this needs to be applied for any such "private
consumer" (although it hopefully goes without saying this way
of consuming libqb outside of it's own playground is hardly
the Right Thing) if portability is important, nonetheless!
* * *
On the address of linker script workaround, there are linkers out there
that do not support the trick, for instance:
- ld.gold from binutils (but it has hardly ever been working with
orphan sections, anyway:
https://sourceware.org/bugzilla/show_bug.cgi?id=22291)
- ancient versions of ld.bfd, e.g. 10+ years old one used as a native
system linker even in the most recent releases of FreeBSD, unless
GCC toolchain is used instead
If these are hit when (because) the compiler has already demonstrated it
supports "section" attribute, the build system configuration is forcibly
stopped, simply to stay conceptually compatible with the prior state in
which the affinity to leverage that feature hasn't been called off
under any circumstances. One is, however, able to achieve exactly
this behaviour with --enable-nosection-fallback switch, but if some
other participants in the logging, possibly linked with a more friendly
linker, do utilize this orphan section, logging may silently break
(another reason to require an explicit sign-off).
Another note, the particular self-check change slightly touched in the
previous commit but otherwise predating this whole effort by far needs
to be modified now once again, this time because linker-script-based
workaround for newer linkers as stated causes the section boundary
symbols to be present regardless if that section is utilized, leading
to a self-inflicted breakage due to these empty section symbols suddenly
winning in the symbol resolution mechanism (previously the empty section
would be dropped incl. the boundary symbols), causing problems down the
line. It also makes this very check self-contained in the same
compilation unit that trigggers it, whereas previously it used to be the
said "arbitrary" winner and things kept silently working just because
failure condition -- empty section -- would be implicitly isolated.
Last but not least, libqb itself needs to be linked with the mentioned
"-Wl,foo.t" equivalent for its own outgoing log messages to be honoured
under all circumstances, which is already achieved with the arrangement
for b. above, and by experiments, further redefinition of those boundary
denoting symbols as weak was necessary so as to make them truly global
within libqb.so proper (at least with binutils 2.29).
* * *
To provide a high-level prioritized overview of what drove the approach:
- PRESERVATION OF BINARY COMPATIBILITY (ABI), which is achieved except
for a single "ABI nongracefulness" I am aware of but that's more
a consequence of slightly incorrect assumptions in the logic of
QB_LOG_INIT_DATA macro function predating this whole affair by
a long shot and which the patchset finally rectifies:
if in the run-time dynamic link, following is combined:
(. libqb, arbitrary variant: pre-/post-fix, binutils < / >= 2.29)
. an "intermediate" library (something that the end executable links
with) triggering QB_LOG_INIT_DATA macro and being built with
pre-fix libqb (and perhaps only with binutils < 2.29)
. end executable using no libqb's logging at all, but being built
with post-fix libqb (and arbitrary binutils < / >= 2.29)
then, unlike when executable is built with pre-fix libqb, the
special callsite data containing section in the ELF structure
of the executable is created + its boundary denoting symbols
defined within, despite the section being empty (did not happen
with pre-fix libqb), and because the symbols defined within the
target program have priority over that of shared libraries in the
symbol resolution fallback scheme, the assertion of QB_LOG_INIT_DATA
of the mentioned intermediate library will actually be evaluating
the inequality of boundaries for the section of the executable(!)
rather than it's own (or whatever higher prio symbols are hit,
presumably only present if the section at that level is non-empty,
basically a generalization of the story so far);
the problem then manifests as unability to run said executable
as it will fail because of the intermediate library inflicted
assertion (sadly with very unhelpful "Assertion `0' failed"
message);
fortunately, there's enough flexibility so as how to fix
this, either should be fine:
. have everything in the executable's library dependency closure
that links against libqb assurably (compile-time) linked with one
variant of libqb only (either all pre-fix or post-fix, mind the
apparent limitation of binutils' versions with the former)
. have the end executable (that does not use logging at all as
discussed precondition) linked using substitution like this:
s/-lqb/-l:libqb.so.0/ (you may need to adapt the number later)
and you may also need to add this CPPFLAG for the executable:
-DQB_KILL_ATTRIBUTE_SECTION
- as high level of isolation of the client space from the linker
(respectively toolchain) subtleties as possible (no new compilation
flags and such required, plus there's no way to hook any dynamic
computational ad-hoc decision when the compilation is about to
happen, anyway), and in turn, versatility is preserved as much as
possible
* * *
Finally, let's have a look how the already well-known test matrix
overview changes as of this commit, but first as a recap,
"X(Y)" denotes "X linked with linker Y":
X(a) .. ld.bfd < 2.29
X(b) .. ld.bfd = 2.29 (+ 2.29.1 and hopefully on)
and here you are (values in <angle brackets> denote non-trivial change
[not mere rewording] introduced as of this commit, in comparison to the
table stated in the preceding commit):
+=========+=========+=========+=========+=========+=========+=========+
#client(x)# libqb(a) usage # libqb(b) usage #
# vvv #---------+---------+---------+---------+---------+---------+
# V # direct | libX(a) : libX(b) # direct | libX(a) : libX(b) #
+=========+=========+=========+=========+=========+=========+=========+
# x = a # OK | OK : <OK> # <OK> | <OK> : <OK> #
# x = b # <OK> | <OK> : <OK> # <OK> | <OK> : <OK> #
+=========+=========+=========+=========+=========+=========+=========+
Everything is green \o/
* * *
Note: as of this fix, it is assumed that the non-green counterpart of
this table in the message for the preceding commit (loosely though[!],
as the occurrence of empty callsite section can no longer be attributed
to something bad going on as of this fix that enforces its presence
unconditionally, whereas it would be suppressed when unused before
with kind linkers, hence some other conditions can be witnessed
especially when QB_LOG_INIT_DATA misused in no-logging context)
doubles as an indicator how will mixing the logging participants wrt.
linker+libqb version work out, when "X(Y)" becomes read as "X linked
with linker Y under additional restriction on libqb version when
compile-time link is performed of the particular part":
X(a) .. ld.bfd < 2.29 OR [arbitrary ld.bfd AND libqb after this fix)
X(b) .. ld.bfd = 2.29 (and likely on) AND libqb up to, but excluding
this fix
* * *
Let's also state some imperfections and loops kept open:
Deficiencies:
* whenever anything is compiled against our install-time-modified
libqb.so so as to force the visibility of the discussed symbols
(or when compiling [with] libqb internally):
> /usr/bin/ld: warning: ../lib/qblog_script.ld contains output sections; did you forget -T?
- not solvable as long as we use the linker script, and there's
hardly any other way not requiring the libqb consumers to adapt
in any aspect
* as already mentioned, lacking compatibility with ld.gold linker and
won't foreseeably be (cf. https://bugzilla.redhat.com/1500898#c7)
- please stick with ld.bfd (i.e. default ld linker), which you
had to do in the past anyway (at least for compiling libqb
itself)
Open questions:
* should we enable attribute((__section__)) for powerpc and other minor
platforms if the feature is proved to be working there as well?
and if/when that's going to happen, we need to figure out the
transition plan to be spread throughout an extended period to keep
the transition smooth -- notably when now-with-callsite-section
clients will get run-time linked with callsite-section-not-a-default
libqb (say upon it's downgrade), and for that, the libqb's support
alone should be enabled year(s) ahead of the actual client space...
* * *
[*] basically GCC's section("SECNAME") __attribute__ annotation of the
global variables + linker described behaviour previously mistakenly
taken for granted
References:
[1] http://oss.clusterlabs.org/pipermail/developers/2017-July/000503.html
[2] https://bugzilla.redhat.com/show_bug.cgi?id=1477354#c2 + comment 8
[3] https://bugzilla.redhat.com/show_bug.cgi?id=1477354#c9
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
Now that the previous commit provides a foundation for what exactly can
go wrong with ld.bfd = 2.29+ linker, let's start reconciling that with
some reasonable assurance that logging is not silently severed, because
realizing the logs are missing is otherwise bound to happen when the
logs are suddently pretty crucial analytical resource :-)
We'll proceed in two steps as detailed.
* * *
As a first step, the table below concludes how the test matrix overview
introduced with a message for the preceding commit (also introducing
log_test_mock.sh runner which got reused here) looks as of this
refreshed sanity check, once QB_LOG_INIT_DATA macro at hand gets applied
(meaning "for non-libqb logging participants" so as not complicate the
matrix further). That macro is nothing triggered directly, it will just
plant a constructor-like function (to be invoked automatically early in
the execution) that will run through the checks (one original and couple
of new ones as of this changeset).
Note that for libqb users, this implies a new link dependency on libdl,
because they may opt-in for refreshed QB_LOG_INIT_DATA sanity check that
calls out to dlopen/dlsym/dladdr directly in case of "attribute section"
being available for the particular platform, and hence immediately needs
those symbols resolved in link time. Hence, add this conditional link
dependency to libqb.pc pkg-config file under Libs variable -- we
actually restore the occurrence of "-ldl" there as it used to be present
until commit 56754d0. While doing so, also move immediate link
dependencies of libqb (if any, currently not but that may be
a regression arising from the cleanup related to the mentioned commit)
represented with the LIBS autoconf variable under Libs.private variable
in libqb.pc, where it belongs per pkg-config documentation.
The promised table follows, but first as a recap, "X(Y)" denotes
"X linked with linker Y":
X(a) .. ld.bfd < 2.29
X(b) .. ld.bfd = 2.29 (and only 2.29)
goes like this (values in <angle brackets> denote non-trivial change
[not mere rewording] introduced as of this commit, in comparison to
the table stated in the preceding commit):
+=========+=========+=========+=========+=========+=========+=========+
#client(x)# libqb(a) usage # libqb(b) usage #
# vvv #---------+---------+---------+---------+---------+---------+
# V # direct | libX(a) : libX(b) # direct | libX(a) : libX(b) #
+=========+=========+=========+=========+=========+=========+=========+
# x = a # OK | OK : BAD[*2] #<BAD[*E]>|<BAD[*F]>:<BAD[*G]>#
# x = b # BAD[*A] | BAD[*B] : BAD[*C] #<BAD[*E]>|<BAD[*F]>:<BAD[*G]>#
+=========+=========+=========+=========+=========+=========+=========+
Woefully, nothing changes if we swap binutils 2.29 for 2.29.1, i.e.,
X(b) .. ld.bfd = 2.29.1, compared to previous state, i.e., the second
table from the previous commit is still applicable for that situation.
The added sanity checks are useful nonetheless, consider for example,
that attribute-section-less libqb is what gets run-time linked to an
attribute-section-full target. The most precise check we could use
-- a custom logger function applied in a self-test scheme -- is not
available at the point the macro-defined function gets invoked, simply
because qb_log_init hasn't been invoked by the time that constructor
gets triggered. However, what we can do is to add a non-trapping
libqb-residing reverse-testing of the client space that (and once it)
voluntarily initiates qb_log_init (delivering abruption all of a sudden
at some unanticipated, as opposed to a well-timed like with
constructors, execution point, seems pretty bad idea + libqb as
a library is a mere helper, not an undertaker :) -- this check then
only announces, via syslog (the only pre-enabled logging target),
the target's logging may be severed.
* * *
Hence, as a promised second step, after incorporating the syslog
change (and extending log_test_mock.sh so as to capture syslog
stream within the container), not much changes with the table above,
i.e., X(b) .. ld.bfd = 2.29:
[*A] in addition, unless QB_LOG_INIT_DATA used on client side,
syslog carries this notice:
"(libqb) log module hasn't observed target chain supplied callsite
section, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly, or target chain built purposefully without
these sections)"
logged by libqb proper
[*C] in addition, unless QB_LOG_INIT_DATA used on interlib side,
syslog carries this notice:
"(libqb) log module hasn't observed target chain supplied callsite
section, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly, or target chain built purposefully without
these sections)"
logged by libqb proper
[*E] in addition, unless QB_LOG_INIT_DATA used on client side,
syslog carries this warning:
"(libqb) log module has observed target chain supplied section
unpopulated, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly)"
logged by libqb proper
[*F] in addition, unless QB_LOG_INIT_DATA used on interlib side,
syslog carries this warning:
"(libqb) log module has observed target chain supplied section
unpopulated, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly)"
logged by libqb proper
[*G] in addition, unless QB_LOG_INIT_DATA used on interlib side,
syslog carries this warning:
"(libqb) log module has observed target chain supplied section
unpopulated, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly)"
logged by libqb proper
but desirably changes with "X(b) .. ld.bfd = 2.29.1" one
(DEP ~ "depends"):
+=========+=========+=========+=========+=========+=========+=========+
#client(x)# libqb(a) usage # libqb(b) usage #
# vvv #---------+---------+---------+---------+---------+---------+
# V # direct | libX(a) : libX(b) # direct | libX(a) : libX(b) #
+=========+=========+=========+=========+=========+=========+=========+
# x = a # OK | OK : DEP[*J] #<BAD[*M]>|<BAD[*M]>:<BAD[*L]>#
# x = b #<DEP[*N]>| DEP[*I] :<DEP[*O]>#<BAD[*M]>|<BAD[*M]>:<BAD[*L]>#
+=========+=========+=========+=========+=========+=========+=========+
* * *
[*1] client logging not working
[*2] interlib logging not working
[*3] both client and interlib logging not working
[*A] boils down to [*1], unless QB_LOG_INIT_DATA used on client side,
which then fails on
"implicit callsite section is populated, otherwise target's build
is at fault, preventing reliable logging"
assertion
[*B] boils down to [*1], unless QB_LOG_INIT_DATA used on interlib side,
which then fails on
"implicit callsite section is populated, otherwise target's build
is at fault, preventing reliable logging"
assertion
[*C] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side,
which then fails on
"implicit callsite section is populated, otherwise target's build
is at fault, preventing reliable logging"
assertion
[*E] boils down to [*1], unless QB_LOG_INIT_DATA used on client side,
which then fails on
"implicit callsite section is self-observable, otherwise target's
and/or libqb's build is at fault, preventing reliable logging"
assertion
[*F] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side,
which then fails on
"libqb's callsite section is populated, otherwise libqb's build is
at fault, preventing reliable logging"
assertion
[*G] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side,
which then fails on
"implicit callsite section is self-observable, otherwise target's
and/or libqb's build is at fault, preventing reliable logging"
assertion
[*I] boils down to [*1], unless QB_LOG_INIT_DATA used on client side,
which makes it, likely through self-reference keepalive (see
below), work OK
[*J] boils down to [*2], unless QB_LOG_INIT_DATA used on interlib side,
which makes it, likely through self-reference keepalive (see
below), work OK
[*K] boils down to [*3]
in addition, syslog carries this notice:
"(libqb) log module hasn't observed target chain supplied callsite
section, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly, or target chain built purposefully without
these sections)"
logged by libqb proper
[*L] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side
(sufficient?), which makes it, likely through self-reference
keepalive (see below), boil down just to [*1];
in addition, syslog carries this notice:
"(libqb) log module hasn't observed target chain supplied callsite
section, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly, or target chain built purposefully without
these sections)"
logged by libqb proper
[*M] boils down to [*1];
in addition, syslog carries this notice:
"(libqb) log module hasn't observed target chain supplied callsite
section, target's and/or libqb's build is at fault, preventing
reliable logging (unless qb_log_init invoked in no-custom-logging
context unexpectedly, or target chain built purposefully without
these sections)"
logged by libqb proper
[*N] boils down to [*M], unless QB_LOG_INIT_DATA used on client side,
which makes it, likely through self-reference keepalive (see
below), work OK
[*O] boils down to [*K], unless QB_LOG_INIT_DATA used on both client
and interlib side, which makes it, likely through self-reference
keepalive (see below), work OK (it's expected that this a mere
composite of situations [*I] and [*J] with consequences as stated)
* * *
Note: the only problematic (i.e. not captured automatically by the
QB_LOG_INIT_DATA macro presumably utilized at every non-libqb logging
system participant in the form of a discrete compilation unit)
combination with 2.29 is the one intersecting at "BAD[*2]" pertaining
"everything but interlib compiled with ld.bfd < 2.29". It would, of
course, be solvable as well, but presumably not in an easy way, and
that use case should not be as frequent.
Takeway: whenever your target (library or client program) actively
utilizes logging (meaning it emits at least a single log message,
otherwise there's an imminent danger of possibly even run-terminating
false positive in the self-check mechanism!),
YOU ARE strongly ENCOURAGED TO USE QB_LOG_INIT_DATA macro function
at (exactly) one of the source code files (presumably the main one)
per respective target's compilation unit.
It will alleviate the hassles possibly caused by downgrading libqb
to the linker-vs-libqb incompatibly compiled one or in similar
circumstances arising merely from the linker behaviour change,
which the current build system/code shake is all about.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
...so as to evaluate use of resources. In particular, the intention
here is to uncover the observable differences between the same logging
code built with callsite section (default when available) and
purposefully (overriding that default by force) without it.
To allow for the latter being applied conveniently, new macro,
QB_KILL_ATTRIBUTE_SECTION, can be defined in the compile-time of the
client code that wishes to opt-out from the callsite section feature.
* * *
Following is a discussion on these differences, sticking with the
logging client code generated with the script defaults, i.e., as it
would be run with these switches:
--callsite-count=3640
--branching-factor=3
--callsites-per-fnc=10
--round-count=1000
and then built twice (as detailed in
tests/functional/log_external/Makefile.am):
* log_callsite_bench_sectionfull
- with callsite section
* log_callsite_bench_sectionless
- without callsite section, imposed with -DQB_KILL_ATTRIBUTE_SECTION
in CPPFLAGS
--> Static size of the executable:
$ size -B log_callsite_bench_section* | tr '\t' ' ' | tr -s ' ' \
| cut --complement -d' ' -f6 | column -t
> text data bss dec filename
> 82761 146180 4 228945 log_callsite_bench_sectionfull
> 190000 588 4 190592 log_callsite_bench_sectionless
We can see that sectionfull is few kB bigger in the object sections
of interest, though the text-data ratio changes considerably, with
code section being cut in half in comparison to sectionless, which
can actually help the code locality (and hence utilization
of CPU caches) in the former case.
--> Dynamic memory/heap operations:
$ valgrind --log-fd=1 ./log_callsite_bench_section{full,less} 2>/dev/null
total heap usage:
- log_callsite_bench_sectionfull:
> 88 allocs, 87 frees, 3,427 bytes allocated[*]
- log_callsite_bench_sectionless:
> 11,894 allocs, 11,893 frees, 486,035 bytes allocated[*]
[*] "32 bytes in 1 blocks still reachable" looks rather as a spurious
warning on the valgrind's side (matter of dynamic linking library)
Apparently, sectionless keeps stirring the heap constantly, with all
the possible downsides associated with that, like hitting the page
faults leading to less timely execution.
--> Run-time efficiency:
$ time ./log_callsite_bench_section{full,less} 2>/dev/null
mean attempts out of 3 consecutive runs:
- log_callsite_bench_sectionfull:
> real 0m1.298s
> user 0m0.965s
> sys 0m0.331s
- log_callsite_bench_sectionless:
> real 0m1.436s
> user 0m1.067s
> sys 0m0.365s
As expected, we can observe sectionfull is slightly faster/more
efficient.
* * *
Based on the above, we can conclude that leveraging the callsite
section for logging as facilitated by the toolchain intrinsics is
beneficial, especially for performance-critical applications (corosync
being the showcase here). Therefore it's desired to struggle for
retaining this nifty trick despite some troubles emerged with recent
binutils releases (starting with 2.29) and the changed behaviour we
relied on so far in respective ld.bfd linkers (as mentioned in
preceding commits). That motive is immediately followed -- well,
judging the impact fairly, actually outclassed -- with the intention
to preserve binary compatibility (incl. continuous library support for
callsite section offloading spread in the existing client space widely
for quite some years already) to the utmost extent possible.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
Based on better understanding how link-time callsite collection works,
put a better description for the macro. Also based on poor user
experience in case that feature does not work well, say because
the linker deliberately changes the previously settled visibility
of the section boundary symbols (happened in ld from binutils-2.29,
fix is forthcoming), tweak the assertion message a bit, together
with an extension of the general intro to point that macro out.
Also play fair, include the <assert.h> header, which this macro
requires.
- use case:
/usr/sbin/pacemakerd --features
- before:
pacemakerd: utils.c:69: common:
Assertion `0' failed
- after:
pacemakerd: utils.c:69: common:
Assertion `"non-empty implicit callsite section" && QB_ATTR_SECTION_START != QB_ATTR_SECTION_STOP' failed.
Restructuring of the assertion inspired by the suggestion of Ferenc
Wágner (for the subsequent commit, actually).
And regarding:
> as a side effect, it can ensure the boundary-denoting symbols for
> the target collection area are kept alive with some otherwise unkind
> linkers
this was actually empirically discovered in one particular combination
with ld.bfd @ binutils 2.29, and extensively with 2.29.1 (placing here
a forward reference for the following commits that elaborate on the
libqb-target cross-combination matrix and the findings).
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
* Fix typo: occur{ -> r}(ed|ing)
* doc: qb-blackbox(8): cosmetic touches
* doc: qb-blackbox(8): add "Portability notes" subsection
* Low: build: git-ignore build-aux/release.mk "overhead" files
This should have been part of ae5138d.
* build: release.mk: ensure checksum file generated even w/o signing
This should have been part of d20e48a.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
* Fix typo: p{rr -> r}ocess
* doc: qbrb.h: several fixes to punctuation
* doc: qbrb.h: reindent example writer's error label as per reader
* doc: qbhdb.h: explain former importance to libqb itself
* doc: ipcc.c: explain why client would timebox recv from server
Also refer to commit d633b4e.
* doc: qblog.h: minor stylistic/doxygen markup cosmetics
* doc: qblog.h: note qb_log_format_set vs. fork interaction wrt. PIDs
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
Under the influence of libxml2 and considering that actual "patch"
information in fact, if significant, ends up encoded in QB_VER_REST,
shift away from convention codified, e.g., by semver.org (not adored
by libqb, anyway) and rename designated PATCH component of the version
to MICRO accordingly.
Note that at this point, after a release without any header-based
versioning present and just a few commits after it was tentatively
introduced, it's a painless change. Once this PATCH nomenclature
is leaked into a full release, there's no way to get rid of it
reasonably...
Mainly as a light-weight alternative to full-blown autoconf/pkg-config
machineries, whereby one can:
* workaround functionality not present in libqb up to 1.0 (inclusive)
- note that this versioning schema is being introduced *after*
1.0.0 release so one cannot tell that version from any older,
but will be able to safely identify any later one (1.0.1+)
and act accordingly
- example:
#if !defined(QB_VER_MAJOR) || ((QB_VER_MAJOR == 1) && (QB_VER_MINOR < 1))
#warning "Feature X not supported"
int do_foo(int arg) { };
#else
int
do_foo(int arg)
{
/* use feature X of libqb */
}
#endif
* make its program report libqb API version it was built with by
emitting QB_VER_STR symbolic string (see tests/print_ver.c for example)
Also added is a print_ver test program to:
* emit how original unparsed version is parsed to particular components
defined in qbconfig.h (QB_VER_{MAJOR,MINOR,PATCH} symbolic integer
constants and QB_VER_REST symbolic string) when being compiled
* emit mentioned QB_VER_STR symbolic string joining the components
back to a single string, plus the components themselves
Resolves: https://github.com/ClusterLabs/libqb/issues/186
Original "qb_log_ctl" interface had to be extended for passing read-only
strings (new parameter), resulting in new "qb_log_ctl2" function, which
is what qb_log_ctl calls into with the new parameter set to NULL.
This ensures backward compatibility.
A new QB_LOG_CONF_IDENT configuration directive for the mentioned
interface is added with a goal to set new internal identifier
that is, notably, used for syslog sink. This allows for switching
the identification without a need to reinitialize logging subsystem,
akin to changing target logging facility.
Also a brand new concept of testing syslog sink in particular is
introduced (finally). During initial trial&error stage, it used
LD_PRELOAD hack but it seems that libtool is sophisticated enough
that no such extra intervention is needed and the desired symbol
resolution Just Works (tm). However, the technique is highly
non-portable (there is even a warning about that from libtool,
which is partially on purpose as the _syslog_override.so should
rather be explicit it is by no mean a regular library) and hence
the syslog tests have to be enabled with explicit
./configure --enable-syslog-tests
rather than possibly break on untested platforms (far too many).
The concept can be extended upon, but initially, just the new
feature is being tested.
Post-review: thanks Chrissie for a suggestion how to deal with
extract-arg-and-forget in a less intrusive way (no defines).
- use semicolons at the end of the description
- use commas when there can be an issue with parsing a long sentence
(and wrap such text at 79 characters if applicable)
Also use the new enum qb_log_target_slot type in for-loops together
with a proper substitute for the literal "0" initializer.
There could be more places that might be type-substituted for this enum
(and hence possibly catch more of an incorrect usage if the compiler
or checker has some notion of enum type narrowing), but leave it as
a possible enhancement for now.
The format strings for QB_HDB_D_FORMAT & QB_HDB_X_FORMAT had no spaces
between "%" and PRIx64. This is allowed, but technically incorrect and
breaks C++11
Signed-off-by: Christine Caulfield <ccaulfie@redhat.com>
If a log message contains a magic character, everything after that marker
will be considered "extended" information. Log targets default to showing
this extended information but qb_log_ctl() may be used to indicate that
the extended information should be stripped instead.
The use case is a log message that has some information of value to administrators
and other information that is only of value for debugging. Two targets can be
configured such that the administrator log is not cluttered with debug information,
but a separate debug log is still available with complete information.
The magic character is the bell (\a). If someone wants to log an actual bell
character, they can do so in the extended information or via a %c format
(only the first bell is magic and only in the format string).
