Frank Ch. Eigler [Tue, 28 Jan 2025 20:32:47 +0000 (15:32 -0500)]
PR32607: undefeat stapconf_HASH caching
Don't try to outsmart make with the autoconf_FOO.c.h ->
stapconf_HASH.h manufacturing steps in the generated module makefile.
If stap knows the result stapconf_*.h was pulled in from the module
cache, there's no need to tell make about the autoconf_* files, or at
least not to activate them in the usual dependency graph. This
restores proper stapconf* caching, absence of which showed sometimes
absurd compilation times on small VMs.
Frank Ch. Eigler [Mon, 20 Jan 2025 14:00:36 +0000 (09:00 -0500)]
systemtap.spec: disable %check by default
This cannot be enabled generally, because build system kernels
may be drastically different from the rpm target distro kernels,
thus the kernel-devel versions may vary or just be unavailable.
The configury is not completed, so machines without
boost/iostreams/device/file_descriptor.hpp just fail to compile
during the build, rather than falling back or exiting during
configure.
Frank Ch. Eigler [Sat, 18 Jan 2025 00:40:56 +0000 (19:40 -0500)]
PR32458: rawhide (6.13) kbuild compatibility
Subject kernel versions introduce changes to the kbuild $(MAKE)
invocations process command line options, and twiddle with working
directories, making our systemtap module Makefile not quite
compatible. Sprinkle $(obj)/ liberally around all the generated
header files (used for autoconf* purposes) and nearby. This appears
to make both >6.13 and <6.13 kbuilds happy.
It appears as though an unnecessarily complicated bit of our
module-building process, invoking a nested $(MAKE) to generate the
STAPCONF_HEADER header file. This stopped working with a fedora
rawhide 6.13-rc thing, leading to frankly mystifying phenomena,
including our own generated Makefile being rewritten (!!!). This
is now simplified.
RHBZ2333874: testsuite: add stap self-test to environment_sanity.exp
Add a "stap -p4 -c stap -e ..." test, so that
% make check RUNTESTFLAGS=environment_sanity.exp
is a good "must-pass" smoke test to verify toolchain installation as
well as basic runtime/kernel compatibility.
Frank Ch. Eigler [Wed, 18 Dec 2024 19:25:19 +0000 (14:25 -0500)]
PR32401: support additional c++ language codes
Extended dwflpp::iterate_over_globals to recognize other c++ language
levels. Extended cast-scope.exp to iterate over several -std=c++FOO
cflags. Tweaked cast-scope.cxx to use argv[1] for string parameter,
to prevent compiler short-circuit optimizations on fixed literal
hello-world.
William Cohen [Mon, 11 Nov 2024 20:32:53 +0000 (15:32 -0500)]
Reduce the emacs dependencies to the bare minimum
The "BuildRequires: emacs" in systemtap.spec was used to pull in the
%{_emacs_sitelispdir} macro. This rpm macro is actually provided by
the emacs-common RPM. Changed the build dependency in systemtap.spec
to emacs-common to minimize the additional RPMs pulled it.
Martin Cermak [Mon, 11 Nov 2024 16:41:53 +0000 (17:41 +0100)]
PR32302: Emit forward decls within tracepoint_derived_probe_group() too
Forward decls added in commit 069def0ae9184 need to actually be
emitted in 2 locations, roughtly corresponding to where
tracepoint_extra_decls() is used.
It would be ideal to put them directly to tracepoint_extra_decls()
but that gives significantly less recognized tracepoints.
The plan is toreview this again though and try to merge the
tracepoint_forward_decls() body into tracepoint_extra_decls()
if possible.
Martin Cermak [Fri, 8 Nov 2024 17:54:10 +0000 (18:54 +0100)]
PR32302: Unrecognized tracepoints
Add forward decls allowing for more kernel headers to be meaninfgully
used for tracepoint identification. With this patch kernel.trace("*")
expands to 3080 tracepoints on 6.12.0-0.rc4.20241022gitc2ee9f594da8.40.fc42
x86_64.
William Cohen [Wed, 6 Nov 2024 21:24:06 +0000 (16:24 -0500)]
Deal with kernels pre-6.12 kernel that relocated unaligned.h
Observed on kernel-6.11.0-28.el10 the <asm/unaligned.h> was moved to
<linux/unaligned.h>. The check was just looking at the kernel version
which was not sufficient on this Linux 6.11 with a backport moving
unaligned.h.
William Cohen [Fri, 25 Oct 2024 17:08:05 +0000 (13:08 -0400)]
Remove obsolete systemtap.base/marker.exp
Kernel markers were removed by git commit fc5377668c3d808e1 in 2009
from the 2.6.32 kernels. At this point marker.exp test just pushes up
the UNTESTED count of the testsuite and does not provide any useful
testing. Removing marker.exp to clean up the testsuite.
William Cohen [Fri, 25 Oct 2024 14:34:14 +0000 (10:34 -0400)]
Use "uname -m" rather than "uname -i" for register_*.exp tests
Several other tests for specific architectures were using "uname -i".
On newer Fedora distributions this returns "unknown" rather than the
hardware platform. On RHEL 8 "uname -i" works. It returns the struct
utsname machine field, the same field that "uname -m" would print.
Switching these register_*.exp tests to use "uname -m".
William Cohen [Thu, 24 Oct 2024 17:27:06 +0000 (13:27 -0400)]
Remove testing of dcookie syscall from systemtap.syscall/arg_write.stp
The dcookie support was removed from the linux 5.12 kernel (git commit be65de6b03aa6). Need to remove testing of dcookie syscall from
arg_write test to allow this test to run on newer kernels. There is
still testing of argument writes on the dir and poll syscalls.
William Cohen [Wed, 9 Oct 2024 01:43:31 +0000 (21:43 -0400)]
Update systemtap.base/cast.stp test to work with newer kernels
Linux kernel commit ec1d281923cf changed the usage field in the
task_struct from atomic_t to refcount_t. The internal fields of the
two types are different. Updated the cast.stp test to use the
appropriate type and field regardless which type is being used.
William Cohen [Wed, 9 Oct 2024 00:54:29 +0000 (20:54 -0400)]
Update systemtap.printf/memory1_module.c to compile with newer kernels
Newer kernels are built with -Wmissing-prototypes enabled. Need to
add prototypes for the exported functions in the module to allow it to
compile. Also select the appropriate struct (struct proc_ops for
newer kernels or struct file_operations for older kernels) so the
stm_fops_cmd argument for proc_create() is the correct type.
William Cohen [Mon, 7 Oct 2024 15:27:10 +0000 (11:27 -0400)]
Address the Linux 6.12 move of the unaligned.h include file
The Linux git commit 5f60d5f6bb moved unaligned.h from asm/unaligned.h
to linux/unaligned.h. Adjusted runtime/sym.c to include the
unaligned.h from the appropriate location.
Sam James [Wed, 21 Aug 2024 06:44:41 +0000 (07:44 +0100)]
PR32107: Drop obsolete -Werror hack
This was added in 2537a8805ce500b4e3474317dcbca4744895326e to workaround
a packaging whoopsie on our side in Gentoo (long before me!). It is now
obsolete and has been for a decade. Clean it up.
William Cohen [Mon, 2 Sep 2024 23:38:47 +0000 (19:38 -0400)]
Adjust the rpc.stp tapset to work with newer kernels
Changes in the tapset made to address:
- $timeout is not available in a number of places (kernel commit c743b4259c3a)
- svc_destroy has $servp in place of $serv
- svc_process is a void function (kernel commit 55fcc7d9159de)
Martin Cermak [Fri, 9 Aug 2024 05:40:43 +0000 (07:40 +0200)]
RHEL-52491: NSS 3.101.0 compatibility
After NSS rebased to 3.101.0, systemtap stopped being able to grant
trust to stap-server: https://issues.redhat.com/browse/RHEL-52491 .
This patch fixes it.
Previous version of this test was painfully verbose. New code
runs stap under "-u" unoptimized mode to compute the backtraces
(to verify non-corruption of memory during those ops), but does
not actually print them any more.
Martin Cermak [Fri, 26 Jul 2024 08:53:59 +0000 (10:53 +0200)]
Generate DWARF for python HelperSDT (refix)
Refix commit 4e8058f38d7f2ec2040d43ddcfddff5217e89423 by letting
CFLAGS propagate to python/setup.py through the Makefile, instead
of passing the flag explicitly. Suggested by Miro Hroncok in
https://issues.redhat.com/browse/RHEL-43481 .
Martin Cermak [Thu, 25 Jul 2024 14:15:53 +0000 (16:15 +0200)]
Generate DWARF for HelperSDT/_HelperSDT*.so
Python Extension Flags Reduction effort [1] dropped GCC -g switch
from the build and linker flags (CFLAGS and LDFLAGS) saved internally
in the Python interpreter for use by distutils and other build systems.
For systemtap this means that the expected debugging symbols weren't
generated for the Helper SDT python module, e.g.:
/usr/lib/debug/usr/lib64/python3.12/site-packages/HelperSDT/_HelperSDT.\
cpython-312-x86_64-linux-gnu.so-5.1-7.el10.x86_64.debug
in systemtap-runtime-python3-debuginfo ... is missing debugging symbols.
To generate the needed debugging symbols, the -g switch needs to be
passed explicitly to the Helper SDT gen script via extra_compile_args.
New sdt-dtrace subpackage for dtrace tool (rhbz#2296275)
This is phase one of the approved self-contained change:
https://fedoraproject.org/wiki/Changes/Separate_dtrace_package
Phase two will be to switch all packages that need the dtrace tool
to the new systemtap-sdt-drace and then we can remove
the requirement from systemtap-sdt-devel package.
William Cohen [Tue, 2 Jul 2024 02:55:43 +0000 (22:55 -0400)]
Update tty.stp tapset for Linux 6.1 and newer
The buildok/tty-detailed.stp test was failing on Linux kernels 6.1 and
newer. The tty.stp tapset was updated to address:
- the n_tty_read() $buf argument becoming $kbuf
(Linux git commit 3b830a9c34d5897be)
- removal of the magic field from tty_struct
(Linux git commit 7a4e0d2c7fb8e28bb)
William Cohen [Mon, 1 Jul 2024 21:06:46 +0000 (17:06 -0400)]
Allow buildok/twentyseven.stp to work on kernels with inlined do_execve()
A earlier systemtap git commit 35ea394c93d3 disabled
buildok/twentyseven.stp probing of inlined versions of do_execve()
because the arguments being used were not available on Linux 4.11
kernels for the inlined version. However, on newer x86_64 Fedora 39
systems the compiler is optimizing the Linux kernel 6.9.7 kernel code
so only the inlined version of do_execve() exists. Loosening the
tewentyseven.stp probe to probe both inlined and non-inlined versions
of do_execve() to allow the test to work on newer systems.
William Cohen [Mon, 1 Jul 2024 20:46:48 +0000 (16:46 -0400)]
Update scheduler.stp tapset to work with newer Linux kernels
The buildok/scheduler-detailed.stp was failing due to a number
of changes in the Linux kernel. The scheduler.stp tapset
has been updated to address those Linux kernel changes:
William Cohen [Fri, 28 Jun 2024 21:06:24 +0000 (17:06 -0400)]
Update nfsderrno.stp to work with Linux 6.10
The NFSERR_OPNOTSUPP error is not a valid NFS described
by any RFC and has been removed from the Linux 6.10
kernels by git commit 939cb14d51a. For the Linux 6.10
kernels just make sure it is set to the value previously
used by earlier kernels for compatibility.
William Cohen [Fri, 28 Jun 2024 18:52:16 +0000 (14:52 -0400)]
Update pmap_agg_overflow.stp to work with Linux 6.10 kernels
The Linux kernel renamed scheduler_tick() to sched_tick()
in git commit 86dd6c04ef9f213e. Adjusted pmap_agg_overflow.stp
to allow the use of either name.
William Cohen [Fri, 28 Jun 2024 15:59:28 +0000 (11:59 -0400)]
Update list of semok tests that do not work with the Dyninst backend
There were a couple of newer semok tests that use kernel specific
probes and/or tapset functions that will not build with the Dyninst
backend. Those are now marked to be KFAIL to accurately reflect
the test results we expect.
William Cohen [Fri, 28 Jun 2024 15:54:15 +0000 (11:54 -0400)]
Update list of buildok tests that do not work with the Dyninst backend
There were a number of newer buildok tests that use kernel specific
probes and/or tapset functions that will not build with the Dyninst
backend. Those are now marked to be KFAIL to accurately reflect
the test results we expect.
William Cohen [Wed, 26 Jun 2024 20:55:23 +0000 (16:55 -0400)]
Update ioblock.stp tapset for Linux 5.11 and newer kernels
The Linux 5.11 kernels removed the $q argument from a number of block
io tracepoints. The tapset has been adjusted to obtain the $q
information via alternative means. The newer kernels have also
removed the bi_phys_segments field from the struct bio (Linux kernel
git commit 14ccb66b3f585). The ioblock.stp tapset now uses the
bi_vcnt field as an approximate representation of that value.
William Cohen [Tue, 25 Jun 2024 02:52:32 +0000 (22:52 -0400)]
Update dev.stp tapset for Linux 6.10
Linux 6.10 changes how the partition number is stored in struct block_device.
Linux kernel git commit 1116b9fa15c097 puts the partition number in the lower
8 bits of the __bd_flags and uses an accessor function to obtain it. The
dev.stp tapset has been adjusted to work with this newer data structure.
William Cohen [Wed, 5 Jun 2024 23:39:28 +0000 (07:39 +0800)]
RISCV registers.stp tapset updated to eliminate -Werror=missing-prototypes error
A number of systemtap scripts running on RISCV would get errors
messages about no previous prototype for
'_stp_regs_get_kernel_stack_nth_addr'. The RISCV registers.stp
tapset has been adjusted to eliminate those errors.
tonyj@suse.de [Tue, 18 Jun 2024 16:52:15 +0000 (09:52 -0700)]
sym.c: fix module_kallsyms_on_each_symbol not exported
module_kallsyms_on_each_symbol has never been exported (as noted by 33fae2d0107f ("This one seems simply like a non-export").
If kallsyms_on_each_symbol is exported (a revert of upstream in our
SLES case) then '!defined(STAPCONF_KALLSYMS_ON_EACH_SYMBOL_EXPORTED)'
is false and the local definition of module_kallsyms_on_each_symbol is
skipped.
This results in:
ERROR: modpost: "module_kallsyms_on_each_symbol" undefined!
Fixes: 33fae2d0107fb6166b4eac3fdffd277829849ab0 Signed-off-by: Tony Jones <tonyj@suse.de>
William Cohen [Wed, 5 Jun 2024 19:38:54 +0000 (15:38 -0400)]
Disable liveness analysis for kernels with CONFIG_MITIGATION_RETPOLINE=y
The upstream linux git commit aefb2f2e619b6c334bc renames
CONFIG_RETPOLINE to CONFIG_MITIGATION_RETPOLINE. The systemtap
code needs to look for CONFIG_MITIGATION_RETPOLINE setting in addition
to CONFIG_RETPOLINE and disable the liveness analysis if either one
is set.
William Cohen [Tue, 4 Jun 2024 13:46:41 +0000 (09:46 -0400)]
Address changes in Linux 6.10 /include/linux/vmalloc.h
Upstream linux kernel git commit 88ae5fb755b0d contains a number of
changes in /include/linux/vmalloc.h that affect vmalloc, vmalloc_node,
and vzalloc_node definitions. These are no longer functions but
macros and cannot be found in the list of symbols exported by the
kernel. Support for vzalloc, vmalloc_node, and vzalloc_node has been
in kernels since Linux 2.6.37. Given that systemtap requires Linux
3.10 or newer there is no longer a need to check for the existence of
these functions or provide local versions of them.
William Cohen [Mon, 3 Jun 2024 19:46:49 +0000 (15:46 -0400)]
Avoid -Werror=empty-body errors from runtime/linux/uprobes-inode.c
Newer linux kernel compiles are being built with -Werror=empty-body.
For some modules generated runtime/linux/uprobes-inode.c is pulled in
and will get error messages like the following:
In file included from /tmp/stapGIM4O9/stap_ded21c54fce18c6570a8930d823aca3a_10928_src.c:2439:
/home/wcohen/systemtap_write/install/share/systemtap/runtime/linux/uprobes-inode.c: In function 'stapiu_change_semaphore_plus':
/home/wcohen/systemtap_write/install/share/systemtap/runtime/linux/uprobes-inode.c:795:5: error: suggest braces around empty body in an 'else' statement [-Werror=empty-body]
795 | ; // already unlocked
| ^
cc1: all warnings being treated as errors
Added "{}" in the appropriate location to indicate to the compiler
that this is intentional.
William Cohen [Mon, 3 Jun 2024 18:40:04 +0000 (14:40 -0400)]
Avoid -Werror=old-style-declaration for stap_probes array in generated kernel modules
With newer linux kernels additional compilers checks are being done
and will get error messages like the following for the generated
module:
/tmp/stapuundLy/stap_2755fca707746de04395c85872aae4b8_1753_src.c:111:1: error: 'static' is not at beginning of declaration [-Werror=old-style-declaration]
111 | } static stap_probes[];
| ^
cc1: all warnings being treated as errors
Tweaked the code generation in translate.cxx to output the static
stap_probes array in a form that is agreeable to newer kernel builds.