1. DTrace Release Notes
  2. DTrace Changelog
  3. 2.0.0-1.2 (Aug 6th, 2020)

2.0.0-1.2 (Aug 6th, 2020)

Second errata of the standalone userspace implementation.

This is a pre-release with limited functionality.

New features:

  • The profile provider has been implemented. Both profile-n and tick-n probes are supported, for probes with default fire rates and user-specified rates/intervals.

  • The trace() action has been updated to provide more consistent output based on the datatype of its argument (signed vs unsigned, and width).

  • The printf() action has been implemented for numeric values.

  • The raise() action has been implemented.

  • Clauses can now be specified with one or more probe specifications, and each probe specification can contain wildcards.

  • Listing probes based on wildcard probe specifications has been implemented.

  • It is now possible to specify the same probe for multiple clauses.

  • Various built-in variables are available: arg0 through arg9 (for probes that provide arguments), curcpu curthread, epid, gid, pid, tid, uid, and timestamp.

  • Expected test outcomes are continually being updated to reflect increasing functionality.

  • The locked-memory limit is raised automatically if it is too small, since BPF has relied on a higher limit.

Bugfixes:

  • Various memory management issues such as memory leak and unsafe memory access operations were fixed.

  • A register allocation leak in predicate handling was fixed.

  • The 'timestamp' built-in variable should yield the same value every time it is used within a specific clause. The cached value should not leak into the next clause execution.

  • Interrupting dtrace using Ctrl-C or sending a signal could leave uprobes and/or kprobes behind. We now ensure that the interrupt handler is set up early enough to be able to provide proper cleanup.

  • There were cases where integer values between INT32_MAX and UINT32_MAX were not processed correctly.

  • The fallback support for /proc/kallsyms did not handle the lack of symbol sizes correctly, making it impossible to map an address to a kernel symbol.

  • int8_t is now always signed, even on platforms where char is unsigned.

Internal changes:

  • The optional predicate for a clause is now compiled as an inline conditional at the beginning of the clause execution.

  • The machine state used during clause execution was allocated on the stack in previous releases. This worked fine when BPF code was executed using the kernel interpreter, but when the JIT BPF engine was used it would result in stack overruns. Now, the machine state is stores in a BPF map value to free up stack space.

  • The creation of uprobes and kprobes as underlying probing mechanism has been deferred until tracing is actually about to start. This means that listing probes (-l) no longer results in uprobes and/or kprobes being created on the system without being used.

  • A significant (but largely invisible) change has been implemented in the D compiler and runtime environment. The compilation of a clause will now merely generate a BPF function. When probe execution is being set up, all probes that are to be part of the tracing session are collected and for each probe a list of associated clauses is created. Finally, when tracing is to commence, a trampoline BPF program is assembled for each probe. Each of the trampoline BPF programs will include calls to the clauses associated with the probe. When the final program for each probe is linked, references to compiled clauses and precompiled BPF utility functions are resolved. At this point, the BPF program for each probe is loaded into the kernel and attached to its probe.