On Fri, 14 Aug 2026 15:55:06 +0200
Tomas Glozar <[email protected]> wrote:

> RTLA is an interface to the kernel osnoise and timerlat tracers, and as
> such, operates on all CPUs seen by the kernel, regardless of any applied
> user space CPU isolation. Currently, it uses libc's get_nprocs_conf()
> function to get the total number of valid (configured) CPUs on the
> system.
> 
> The number reported by the function may vary depending on both the
> libc implementation and the environment. This can cause discrepancies,
> including:
> 
> - RTLA crashing because of out-of-bounds access and/or memory
>   corruption, as a result of reading kernel events with cpu >= nr_cpus
>   and using it to index an nr_cpus-length buffer.
> - nr_cpus mismatching the return value of libbpf_num_possible_cpus(),
>   which is used in the per-CPU map size check, leading RTLA to fail to
>   read BPF data.
> 
> To address both issues, this patchset switches RTLA to unconditionally
> use the CPU count exposed through sysfs
> (/sys/devices/system/cpu/possible) - the same interface through which
> the tracers are accessed. As several places in RTLA assume the CPU count
> correlates with the maximum CPU (pre-existing limitation), the new
> mechanism also rejects possible CPU lists with holes or starting with
> non-zero CPU.
> 
> As a safety measure, trace events are further validated and rejected
> if their cpu field is greater or equal to nr_cpus. nr_cpus is
> additionally validated for equality against osnoise's "all" cpu setting.
> This is done to catch environments where sysfs is virtualized and
> reports a different number of CPUs than the kernel.
> 
> Note: As an alternative approach - reading only the osnoise cpu "all"
> setting and using it for nr_cpus - was considered. While it would make
> patch 3 unnecessary and make RTLA compatible with environments with
> virtualized sysfs, it would require separate handling of userspace and
> kernel CPU number, as nr_cpus is also used to guard -H/--housekeeping
> option argument. That does not appear to me to be worth it just to cover
> for a very rare corner case; if such use case appears in the future, RTLA
> can always be fixed.
> 
> Tomas Glozar (5):
>   rtla: Replace get_nprocs_conf() with sysfs possible cpus
>   rtla: Discard trace entries with cpu >= nr_cpus
>   rtla: Abort on nr_cpus mismatch with tracer
>   rtla/tests: Add unit test for cpu_list_iterate()
>   rtla/tests: Add unit test for get_max_cpu_from_list()
> 
>  tools/tracing/rtla/src/common.c        |  29 +++-
>  tools/tracing/rtla/src/common.h        |   3 +-
>  tools/tracing/rtla/src/osnoise.c       |  70 +++++++++-
>  tools/tracing/rtla/src/osnoise.h       |   2 +
>  tools/tracing/rtla/src/osnoise_hist.c  |   1 +
>  tools/tracing/rtla/src/osnoise_top.c   |   1 +
>  tools/tracing/rtla/src/timerlat_hist.c |   1 +
>  tools/tracing/rtla/src/timerlat_top.c  |   1 +
>  tools/tracing/rtla/src/trace.c         |   7 +
>  tools/tracing/rtla/src/trace.h         |   1 +
>  tools/tracing/rtla/src/utils.c         | 181 +++++++++++++++++++++----
>  tools/tracing/rtla/src/utils.h         |   6 +
>  tools/tracing/rtla/tests/unit/utils.c  |  71 ++++++++++
>  13 files changed, 342 insertions(+), 32 deletions(-)
> 

Note, the merge window will likely open next week (Sunday) (and of course,
I'll be traveling). If you want this in the next release, please send a pull
request today. Otherwise it will likely not be able to go into Linus's tree
until 7.4.

-- Steve


Reply via email to