Release Notes for DTrace 2.0.0 1.14

2.0.0-1.14 (Mar 5th, 2024)

Fourteenth errata of the standalone userspace implementation.

This is a pre-release with limited functionality.

The io provider has been implemented.
The print() action has been implemented.
The link_ntop() subroutine has been implemented.
The cleanpath() subroutine has been implemented.
The d_path() subroutine has been implemented to always return "<unknown>". This is needed to ensure that the io and procfs translators compile.
The -xcpu option has been implemented.
The -xaggpercpu option has been implemented.
The -xlockmem option has been improved. The limit is set before retrieving probe info, and the default behavior is now "unlimited" (meaning most users will not have to worry about this option).
The pid provider now supports offset-based probe names.
Aggregations of stacks are now supported.
The retrieval of rawtp argument information has been improved.
It is now possible to delete an element in an associative array by assigning a literal 0 to it, regardless of the element datatype.
The lexer has been improved to support module names that start with a numeral so that they can be used (e.g. 9p`v9fs_remove).
A basic configure script has been added to facilitate building and packaging in a variety of distributions.
USDT probe information maintained by dtprobed is now stored under /run to ensure it can survive daemon restarts.

Drop counter handling is fixed for local-only updates.
Dedicated space has been introduced for call stacks, so that stackdepth and temporary strings will not overwrite one another.
dt_tp_event_info() has been corrected so as not to overrun its buffer.
Compilation of BPF code that uses BPF helpers now uses the bpf_helpers.h header file from libbpf-dev[el] instead of the (deprecated) bpf-helpers.h header file that the gcc BPF cross provided.
Due to the need to support DTrace on older kernels, BPF source code files are now compiled using -mcpu=v3 to ensure that the object code is acceptable to the BPF verifier in older kernels.
When a dtrace instance would trigger the END probe to be processed, any and all other dtrace instances on the system would have their END probe fire as well because the dtrace provider trampolines were not validating the tgid of the task triggering the probe.
The initialization of the cpuinfo BPF map could cause a buffer overrun on systems with non-sequential online CPU ids.
On kernels that support preemptive BPF program execution, probe data could get corrupted. As a temporary fix, concurrent BPF program execution for DTrace probes is blocked.

Code has been restructured to better support SDT-based providers. While such providers (lockstat, io, etc.) used to be based on static probes in the kernel source, they are now implemented with fbt, rawtp, and even syscall probes. Probe trampolines can become involved. Changes, notably in cg, better support these providers. Also, the underlying probes are using rawtp more rather than relying just on fbt.
Support for compilation in older environments (esp. older compilers) has been improved.
There have been a number of build improvements, especially for cross compilation and to build with upstream kernels.
A bunch of code to parse strings has been removed, relying instead on flex for this support.
The creation and deletion of USDT probes has moved from dtprobed to dtrace.
The dtprobed now uses presets for daemon restarting.

Test dependence on tick-* probes has further been reduced. The tick-* probes can behave poorly on some kernels, depending on how their timers subsystem is configured (CONFIG*_HZ*). Reducing this dependence has gone on over multiple releases to improve the robustness of these tests.
Fix err.* tests that force XFAIL to report so correctly.
Skip lockstat testing prior to 5.10.
Fix the use of syscall::execve:entry args[1][?], since there are two levels of dereferencing userspace addresses, requiring two copyin*().