(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>
...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>
These are for quick manual sanity checking, assuming the target audience
-- maintainers -- are clear on the context of use and the purpose
(perhaps with the help of static files for comparison and/or additional
checking harness, usually available through "make check", but not to be
confused with regular unit + broader tests). These test are meant to be
compiled on demand only, not during the standard building routine, for
which a trick leveraging GNUmakefile-Makefile precedence with GNU make
was devised (GNU make/gmake already required by configure script for
other reasons [some pattern-based matching not available with FreeBSD's
default "make", IIRC], so this introduces no new build dependency).
The respective new tests are meant to simulate logging variants in two
different library consumption models:
a. regular: linking against system-wide library
b. developmental: consuming library from a local sub-checkout tree,
using libtool conventions and hence attaching the
library through libqb.la intermediate library
descriptor of libtool
and between up to three possibly affected logging system participants
(discrete compilation units):
1. libqb itself will emit log messages in boundary conditions or
for tracing purposes
2. client program that consumes libqb's logging API directly
3. ditto, but the client program furthermore links with a library
(referred to as "interlib") that itself exercises the logging
API (it's also linked with libqb) -- through induction, this
should cover whole class of N interlib cases
Especially the latter perspective makes for a test matrix to possibly
(hopefully) demonstrate a fix allowing to cope with the changed
behaviour of ld from binutils 2.29+ wrt. boundaries denoting symbols for
a (custom) orphan section that are no longer externally visible. Such
commit is in the pipeline...
Developmental consumption model (a.) is now also tested automatically
in Travis CI runs and as a part of %check within upstream-suggested
libqb.spec for RPM packaging, whereas the regular one (b.) serves as
a building block for new log_test_mock.sh runner of said test matrix
-- it iterates through all the possible permutations of linker-imposed
implicit visibility of mentioned symbols between various affected
link participants all making use of logging (see 1. - 3. above) so as
to demonstrate A/ the impact of the problem (see table below), and
subsequently B/ that the fix is effective in all these situations
(updated table will be provided as well) once it lands. This script
also allows convoluting the test matrix further, notably with on-demand
defusing the self-checks based on QB_LOG_INIT_DATA macro, which is
of significance as demonstrated below (and will become even more
important with upcoming patches in this series).
* * *
Current state for such matrix, in which participants 1. - 3. map like:
1. ~ libqb(Y)
2. ~ "direct"
3. ~ libX(Y) [a.k.a. interlib]
and where "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:
+=========+=========+=========+=========+=========+=========+=========+
#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[*1] | BAD[*D] : BAD[*3] #
# x = b # BAD[*A] | BAD[*B] : BAD[*C] # BAD[*1] | BAD[*C] : BAD[*3] #
+=========+=========+=========+=========+=========+=========+=========+
whereas if we swap 2.29 for 2.29.1, i.e., X(b) .. ld.bfd = 2.29.1, we
can observe a somewhat simpler story (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[*1] | BAD[*1] : BAD[*L] #
# x = b # DEP[*I] | DEP[*I] : DEP[*K] # BAD[*1] | BAD[*1] : 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 fails on 'implicit callsite section is populated' assertion
[*B] boils down to [*1], unless QB_LOG_INIT_DATA used on interlib side,
which fails on 'implicit callsite section is populated' assertion
[*C] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side,
which fails on 'implicit callsite section is populated' assertion
[*D] boils down to [*3], unless QB_LOG_INIT_DATA used on interlib side,
which makes it boil down just to [*1] (hypothesis: mere internal
self-reference to the section's boundary symbols makes them
overcome some kind of symbol garbage collection at the linkage
stage, so they are exposed even they wouldn't be otherwise as
demonstrated with the initial, plain case of [*3])
[*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], 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)
[*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]
* * *
Note: as observed with [*D] case (libqb linked with ld.bfd < 2.29
whereas interlib and its client linked with ld.bfd = 2.29), the exact
availability of a working logging doesn't depend solely on the linkers
in question, but generally (further investigation out of scope) the
conclusion is that when 2.29 ld.bfd is involved somewhere in the chain
of logging-related discrete compilation units, also (self-)referencing
of the section's boundary denoting symbols is a deciding factor whether
particular logging source will be honored. This may be a result of
some internal linkage garbage collection mechanisms involved.
Anyway, it is supposed that the fix to broken-by-linkage logging can be
proclaimed complete once all combinations pass barring QB_LOG_INIT_DATA
usage (incurring the mentioned active referential use of the symbols),
along with a spin using it everywhere for good measure.
For another level of the analysis depth, one can further play with
combinations of -n{sc,cl,il} options (explained upon -h switch) to
log_test_mock.sh (taking an oracle, this is added mostly to justify
the upcoming self-check test change because linker-script-based
workaround for newer linkers will cause 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).
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>
Various "configure" commits by wferi recently introduced new
compat/custom m4 macro files in m4 directory, which itself was,
so far, assumed ephemeral (not strictly needed for reproducing
the build successfully, i.e., bits that can be completely purged
when cutting down the project files to the bone). Apparently,
this assumption no longer holds so several places need to be
adapted.
Amonst others, m4 directory no longer needs to be reinsured in
autogen.sh, and special care must be taken with .gitignore
and maintainer-clean-local target of the main Makafile.
Signed-off-by: Jan Pokorný <jpokorny@redhat.com>
My first attempt at this broke "make rpm" - oops.
It seems that we need a common extension for tests to get the
dependancies to work.
Signed-off-by: Angus Salkeld <asalkeld@redhat.com>
This ringbuffer is usable across processes.
the point is to use this for IPC to provide async
connections from client to server, but with inherient
flow control.
This still needs a bit of clean up, but committing now
for feedback and as it is quite functional.
Signed-off-by: Angus Salkeld <asalkeld@redhat.com>
