Previous approach with passing the section contents as a command line
option proven to be error-prone, as it hits the command line option size
limit on some kmods. Move the reading and processing of the section
contents inside the awk script itself; it also saves a bit (around 3—4%)
of time on the kernels with indirect __crc_* symbols:
$ for i in ./lib/modules/*; do \
echo "====== $i ====="; \
diff -u <(find $i | ./find-provides.ksyms.old) <(find $i | ./find-provides.ksyms.new); \
echo -n "old: "; find $i | time ./find-provides.ksyms.old > /dev/null; \
echo -n "new: "; find $i | time ./find-provides.ksyms.new > /dev/null; \
done
====== ./lib/modules/4.18.0-372.57.1.el8_6.s390x =====
old: ./find-provides.ksyms.old > /dev/null 5.38s user 4.66s system 150% cpu 6.693 total
new: ./find-provides.ksyms.new > /dev/null 5.20s user 4.52s system 149% cpu 6.484 total
====== ./lib/modules/4.18.0-372.57.1.el8_6.x86_64 =====
old: ./find-provides.ksyms.old > /dev/null 7.85s user 6.34s system 143% cpu 9.864 total
new: ./find-provides.ksyms.new > /dev/null 7.75s user 6.32s system 143% cpu 9.809 total
====== ./lib/modules/5.14.0-284.15.1.el9_2.s390x =====
old: ./find-provides.ksyms.old > /dev/null 6.19s user 4.68s system 144% cpu 7.506 total
new: ./find-provides.ksyms.new > /dev/null 5.93s user 4.46s system 143% cpu 7.219 total
====== ./lib/modules/5.14.0-284.15.1.el9_2.x86_64 =====
old: ./find-provides.ksyms.old > /dev/null 8.47s user 6.71s system 144% cpu 10.523 total
new: ./find-provides.ksyms.new > /dev/null 8.44s user 6.59s system 144% cpu 10.435 total
====== ./lib/modules/6.4.0-0.rc1.20230511git80e62bc8487b.19.eln126.s390x =====
old: ./find-provides.ksyms.old > /dev/null 5.21s user 4.09s system 145% cpu 6.383 total
new: ./find-provides.ksyms.new > /dev/null 5.04s user 3.85s system 145% cpu 6.120 total
====== ./lib/modules/6.4.0-0.rc1.20230511git80e62bc8487b.19.eln126.x86_64 =====
old: ./find-provides.ksyms.old > /dev/null 8.68s user 5.77s system 139% cpu 10.326 total
new: ./find-provides.ksyms.new > /dev/null 8.47s user 5.57s system 139% cpu 10.067 total
* find-provides.ksyms: Add assign to revbytes the result of endianness
chack test instead of the code snippet; do not generate SECTDATA
variable; do not pass SECTDATA to the awk script; pass module
and revbytes to the awk script; retrieve the section data
into the sectdata variable and perform the byte re-shuffling
in accordance to the revbytes value in the BEGIN section of the awk
script.
* find-requires.ksyms (all_provides): Likewise.
Fixes: cd7e9e8a2f "find-provides.ksyms, find-requires.ksyms: rewrite indirect CRC parsing"
Resolves: #2135901
References: https://bugzilla.redhat.com/show_bug.cgi?id=2178935
Co-Authored-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Currently, it just matches the 2nd to 5th fields with awk, which leads
to occasional capturing of some garbage at the end of the section data
and minor issues when supplying it to the symbol extracting awk script:
awk: cmd. line:1: (FILENAME=- FNR=19) fatal: argument index with `$' must be > 0
awk: warning: escape sequence `\.' treated as plain `.'
awk: cmd. line:1: (FILENAME=- FNR=9) fatal: not enough arguments to satisfy format string
`78b624dbc125415351be965ex.$..%ASQ..^'
^ ran out for this one
While at it, also eliminate grep and sed calls (with the pair of awk and
sed being the most offending).
* find-provides.ksyms: Generate part of the awk script that performs
the byte reversal based on the endianness in the readelf -h output,
inject it into the awk script; use substr for matching parts
of the section data instead of fields.
* find-requires.ksyms (all_provides): Likewise.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
To bring it more in line with the way it is implemented
in find-provides.ksyms and avoid the inconsistencies prodiced by missing
the environment setting.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
To avoid all kinds of fun that might stem from interpreting the user-supplied
input.
* find-provides.ksyms: Supply -r option to the read call.
* find-requires.ksyms (all_provides): Likewise.
Fixes: cd7e9e8a2f "find-provides.ksyms, find-requires.ksyms: rewrite indirect CRC parsing"
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Instead of ignoring the section name in the objdump -t output, section
name has to be checked in the awk script to produce only the meaningful
output; hit when both __kcrctab and __kcrctab_gpl are present.
* find-requires.ksyms (all_provides): Pass $sectname to the awk script,
check that the relevant field of the input matches it.
* find-provides.ksyms: Likewise.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Apparently, these are local in the recent kernels.
* find-provides.ksyms (all_provides): Filter local __crc_* symbols
in additional to global in the objdump -t call output.
* find-requires.ksyms: Likewise.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Check for symvers.xz support in addition to symvers.gz,
due to unnanounced unilateral change[1][2] in the kernel packaging.
[1] 2644665657
[2] 2db77a072a
* kabi.attr: Add xz extension to the file matching pattern
* kabi.sh: Add xz extension to the matching pattern; determine catenation
program based on extension.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
References: https://bugzilla.redhat.com/show_bug.cgi?id=2192895
[^set] is not correct in glob (it's regexp construct). [!set] is.
sed: -n is not necessary. -e too. manpage recommends using -E instead of -r.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
The time to run on tests/* data decreased from 34.517s to 30.201s.
Do not allow combining of any glob meta characters, either in first
or second line.
The code is still buggy, it may attempt to combine
modalias(abc[def]123)
modalias(abc[dXf]123)
into
modalias(abc[d[eX]f]123)
(presumably there won't be any such cases in real modules, but still).
Eliminate "| head -n1" - do it with bash string ops.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
On my machine, out of 2604 modules, this reduces number of
"provides" items by ~800 (to 12689 items).
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
This creates a subshell: ... && ( [ -z "$pos" ] || [ $n = $pos ] )
For example, processing ath9k.ko needed ~80 forks here.
Rather expensive.
This should do the same: ... && [ -z "$pos" -o "$n" = "$pos" ]
While at it, explain algorithm in comments.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
While (in this case) their effect is the same,
many more people know that "local" makes variables function-local,
than that "declare" in bash does the same. I didn't know.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
basename(): return one value, not two (principle of least confusion).
strip_compress_sfx(): ditto
strip_compress_sfx(): fix it, it had several problems:
(a) incorrect escaping of . (period) ... masked by incorrect use of backslash
(b) lua has no | "or" pattern
Use simpler gsub patterns / replacements, presumably faster (did not measure).
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Starting with RHEL 9.2 and Fedora 35, debuginfod is enabled by default,
which leads to significant startup times of various tools
from the elfutils package (namely, readelf and objdump).
Unset DEBUGINFOD_URLS to avoid it.
* find-provides.ksyms (DEBUGINFOD_URLS): Unset.
* find-requires.ksyms (DEBUGINFOD_URLS): Ditto.
* kernel-srpm-macros.spec (Release): Bump to 18.
(%changelog): Add a record.
Suggested-by: Panu Matilainen <pmatilai@redhat.com>
Co-Authored-by: Denys Vlasenko <dvlasenk@redhat.com>
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Linux commit v5.19-rc1~139^2~2 ("kbuild: link symbol CRCs at final link,
removing CONFIG_MODULE_REL_CRCS") has broken the assumption aobut
the placement of non-absolute CRC symbols in .rodata (and also affects
the architectures that used to have these symbols as absolute); rewrite
the parsing by utilising "objdump -t" output to figure out
the section(s) where the __crc_* symbols are stored and process
it instead of the hard-coded ".rodata" section. The change also speeds
up the processing a bit, around 33% on synthetic tests:
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.ppc64le -name "*.ko.xz" | find-provides.ksyms.old > /dev/null 2> /dev/null'
sh -c 10.36s user 5.58s system 137% cpu 11.613 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.ppc64le -name "*.ko.xz" | find-provides.ksyms.new > /dev/null 2> /dev/null'
sh -c 7.82s user 4.59s system 142% cpu 8.686 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.ppc64le -name "*.ko.xz" -exec sh -c "echo {} | find-provides.ksyms.old" \; > /dev/null 2> /dev/null'
sh -c 11.85s user 6.76s system 129% cpu 14.318 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.ppc64le -name "*.ko.xz" -exec sh -c "echo {} | find-provides.ksyms.new" \; > /dev/null 2> /dev/null'
sh -c 8.91s user 5.51s system 135% cpu 10.647 total
* find-provides.ksyms: Process "objdump -t" output to get the list
of sections where __crc_* symbol contents are placed; retrieve
each one with "readelf -R" and supply it to an awk script that cuts
the required part of it for each __crc_* symbol in that section.
* find-requires.ksyms (all_provides): Likewise.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Resolves: #2135047
Commit binutils-2_33~1385[1] has changed (and binutils-2_35~1768[2]
has not reverted) the calculated type for symbols in read-write .rodata section
from 'R' to 'D' on architectures where CONFIG_MODULE_REL_CRCS is enabled
(namely, ppc64* and s390*), since, apparently, many kernel modules have it indeed
read-write. Work around it by matching both D and R as possible __crc_* symbol
type.
[1] https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;h=a288c270991d
[2] https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;h=49d9fd42acef
* find-provides.ksyms: Match D as possible __crc_* symbol type in
addition to R.
* find-requires.ksyms: Likewise.
* kernel-srpm-macros.spec (Release): Bump to 11%{?dist}.
(%changelog): Mention the change.
Resolves: #2055464
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Avoid running sed on the whole "nm" output and also avoid "sed | awk"
pipe in favor of a single awk call. Overall, this gives around 20%
speedup on some quick synthetic tests:
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.x86_64 -name "*.ko.xz" | find-provides.ksyms.old > /dev/null 2> /dev/null'
sh -c 14.20s user 8.93s system 144% cpu 16.014 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.x86_64 -name "*.ko.xz" | find-provides.ksyms.new > /dev/null 2> /dev/null'
sh -c 12.01s user 7.46s system 143% cpu 13.567 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.x86_64 -name "*.ko.xz" -exec sh -c "echo {} | find-provides.ksyms.old" \; > /dev/null 2> /dev/null'
sh -c 16.31s user 10.77s system 134% cpu 20.092 total
$ time sh -c 'find ./lib/modules/5.14.0-258.el9.x86_64 -name "*.ko.xz" -exec sh -c "echo {} | find-provides.ksyms.new" \; > /dev/null 2> /dev/null'
sh -c 13.95s user 8.92s system 135% cpu 16.836 total
* find-provides.ksyms: Check presence of absolute __crc_* symbols with
"grep -q" exit code and not presence of sed output; rewrite awk script to match
the __crc_* symbols instead of preprocessing the nm output with sed.
* find-requires.ksyms: Likewise.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Resolves: #2135047
This makes the sort order a bit more predictable across various
environments and also speeds up grep/sed/sort a bit as they no longer
need to deal with non-trivial collations and multibyte sequences
that may appear when some garbage is being parsed accidentally.
* find-provides.ksyms: Add "export LC_ALL=C", remove "LC_ALL=C" from the
specific calls.
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
Resolves: #2135047
Add conflicts of kernel-srpm-macros with kernel-rpm-macros < 185-9
as macros.kmp, kmodtool, and rpmsort were moved from the latter
to the former.
* kernel-srpm-macros.spec (Conflicts): Add kernel-rpm-macros < 185-9.
(%changelog): New entry.
Resolves: #2002887
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
* rpmsort: Chnge the FSF address from "59 Temple Place, Suite 330,
Boston, MA 02111-1307 USA" to "51 Franklin Street, Fifth Floor,
Boston, MA 02110-1301 USA".
Resolves: #2002887
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
* modalias.prov: Replace everyithing that is not "0-9a-zA-Z.:"
in a kernel module version with underscores; replace everything
that is not "][0-9a-zA-Z.:*_?/-" in a kernel module alias with
underscores.
Resolves: #2002887
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
* modalias.prov: Chnge interpreter from "/bin/sh" to "/bin/bash" because
it uses bashisms; add "-efu" flags to keep things in check; initialise
is_kernel_package variable; check for presence of /sbin/modinfo, sed,
and sort.
(combine_modaliases): Initialise the "variants" and "pos" variables.
Resolves: #2002887
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
After split of the kmods into a separate packages (like kernel-modules
and kernel-modules-extra), kernel() provides for the inbox kmods are generated
for incorrect package, as they were handled by kabi.sh that uses symvers
as a basis for the dependency list. Stop using it for kmod dependencies
(but continue using it for the symbols provided by vmlinux itself) and
employ find-provides.sh for that purpose.
* kabi.sh: Filter only those symbols that are exported by vmlinux.
* find-provides.ksyms: Generate tags with the "kernel" prefix for kernel
modules inside /lib/modules/[1-9][^/]*/kernel.
Resolves: #1942563Resolves: #1975927Resolves: #2002887
Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
