The London Perl and Raku Workshop takes place on 26th Oct 2024. If your company depends on Perl, please consider sponsoring and/or attending.


Devel::RingBuffer - Shared memory ring buffers for Perl scripts diagnosis/debug


    #    create ringbuffer
        use Devel::RingBuffer;
        use Devel::RingBuffer::TieInt;

        my $ringbuf = Devel::RingBuffer->new(
                File => 'somefile.trace',
                Rings => 20,
                Slots => 20,
                SlotSize => 300,
                MessageSize => 256,
                GlobalSize => 24 * 1024,
                StopOnCreate => 0,
                TraceOnCreate => 1) || die "Can't create a ring buffer.";

        my $ring = $ringbuf->allocate();


Provides shared memory structures (using memory mapped files via IPC::Mmap) to be used by diagnostic and debugger applications for Perl scripts (see Devel::STrace). Using XS/C code to maximize performance, creates a set of ringbuffers with a configurable number of slots. Each slot includes a field for a linenumber, a timestamp, and a fully qualified subroutine name. Each ring buffer also includes additional headers and fields to support diagnostic interfaces, e.g., watched expressions, command/reponse interfaces to the monitored applications, etc.


AUT - Application Under Test; the application started with -d:STrace
Monitor - monitoring application, i.e., the debugger view

Ring Buffer Structure

The ring buffer structure is configured externally using the following set of environment variables:


the number of ring buffers (default 20). In practice, each thread of each process of the AUT allocates its own buffer from this pool, so this needs to be large enough to handle the maximum number of concurrent processes/threads.


the number of slots per ring buffer (default 10)


Name of the file (or the Win32 namespace). Default is the script name, sans any file qualifiers, with the PID and timestamp as the file qualifier, created in the /tmp directory, e.g.,


size in bytes of the global message area aka GMA (default 16K). This area is a large buffer used for communicating large data between the monitor and AUT, and is shared by all threads of the AUT.


size in bytes of the per-thread command/response message area (default 256 bytes). Each ring buffer includes a 4 byte command/response tag, and an associated message area where control information can be exchanged between the AUT and the monitor. This area is used, e.g., to send breakpoint information from a monitor to the AUT.


Sets slot size. Default size is 200 bytes, plus the integer linenumber and double precision timestamp header.


Sets the global Stop On Create flag. This flag causes all newly created threads (including root threads of new processes) to have their $DB::signal flag set, thus causing them to stop in DB::DB() and wait for a command. Default is off.


Sets the global Trace On Create flag. This flag causes all newly created threads (including root threads of new processes) to have their $DB::trace flag set, thus causing them to enter DB::DB() for purposes of tracing current statement execution.

Note that monitored applications with large numbers of threads or processes, and that use large msgarea fields and/or large numbers of ring slots, can lead to very large memory mappings, which may cause memory management issues.

The following C type definitions represent the structure of the memory mapped ring buffer file (refer to ringbuffer.h for precise definitions):

    typedef struct {
        int single;                  /* tied to $DB::single (global) */
        int msgarea_sz;              /* size of RingBuffer.msgarea */
        int max_buffers;             /* max number of buffers available */
        int slots;                   /* number of slots per buffer */
        int slot_sz;                 /* size of each slot (plus linenumber and timestamp header) */
        int stop_on_create;          /* 1 => new threads created with ring.signal = 1 */
        int trace_on_create;         /* 1 => new threads created with ring.trace = 1 */
        int global_sz;               /* size of RingBuffers.global_buffer */
                int globmsg_sz;              /* size of current global msg contents */
                char global_buffer[];        /* global message buffer (large, >16K) */
        char free_map[];             /* booleans to indicate if the
                                        buffer of same index is free */
        ring_buffer_t rings[];       /* the ringbuffers */
    } ring_buffers_t;

A global file lock (or semaphore on Win32 platforms) is used control access to some members of the memory mapped file, specifically, the global message area (see below), the free ring map, and during initialization of the entire file. In addition, each process in the AUT uses an additional threads::shared locking variable to control thread-level access.

Note that any non-char fields in these structures are pack()'ed in platform specific format, and must be unpack()'d when read from the structures.

  • single is intended to be tie'd to $DB::single; this provides a global (i.e., all threads of all processes) single-step flag that can be set by the Monitor, such that all threads will enter DB::DB() for each executable statement.

  • msgarea_sz indicates the size of the per-thread Monitor <=> AUT command/response message area. The message area consists of an integer command-ready flag, a 4 byte command field, an integer message length, and the message area of size $ENV{DEVEL_RINGBUF_MSGSZ} (default 256 bytes).

  • maxbuffers indicates the number of ring buffers in the file, as set by $ENV{DEVEL_RINGBUF_BUFFERS} (default 20).

  • slots indicates the number of ring slots per buffer, as set by $ENV{DEVEL_RINGBUF_SLOTS} (default 10). The slots are used to store the current call stack, and track the linenumber/timestamp of the last execution within each subroutine on the stack.

  • slot_sz specifies the size of each slot in bytes, plus the integer linenumber and double precision timestamp; as set by $ENV{DEVEL_RINGBUF_SLOTSZ} (default 200). For monitor/debug applications which supply significant additional information for each logged slot entry (e.g., including the complete argument list of a subroutine call), the default size may be insufficient.

  • stop_on_create indicates that newly created threads should enter single step mode by setting the $DB::signal flag of their ring buffer. stop_on_create permits the Monitor to trap and initialize any per-thread context on thread creation; if not set, newly created threads would simply run until (a) the Monitor detected their new ring and (b) the Monitor set any of the single, trace, or signal flags. May be initialized by $ENV{DEVEL_RINGBUF_SOC}; default is 0 (off).

  • trace_on_create indicates that newly created threads should enter trace mode by setting the $DB::trace flag of their ring buffer. trace_on_create permits the Monitor to assure that all threads are traced immediately upon creation. If not set, newly created threads will not be traced until the Monitor detects their new ring and explicitly sets its trace flag. May be initialized by $ENV{DEVEL_RINGBUF_TOC}; default is 0(off).

  • global_sz indicates the size of the global message area shared by all threads. The intent is to provide an area for very large messages (e.g., transfering a large string to the Monitor). The size of the area is set by $ENV{DEVEL_RINGBUF_GLOBALSZ} (default 16Kbytes). Access to the global message area is controlled by the same file level and thread-level locks as the entire ring buffer file.

  • globmsg_sz indicates the size of the current contents of the global message area. Note that some commands may cause chaining of the contents of the global message area to accumulate very large messages to be "chunked".

  • global_buffer is the global message area itself, of size specified by global_sz.

  • free_map is a map indicating if the ring buffer of the corresponding index is free (== 1) or in-use (== 0). AUT's must acquire the global file lock and the process-local thread lock before manipulating the free map.

  • rings is the the set of per-thread ring buffers (see next section).

Per-thread Ring Buffer

Each thread of each process in the AUT allocates a ring buffer on its first pass through either DB::DB() or DB::sub() to communicate its execution state, and perform per-thread interactions with the Monitor:

    typedef struct {
        int pid;                    /* pid of slot buffer owner */
        int tid;                    /* tid of slot buffer owner */
        int currSlot;               /* current slot */
        int depth;                  /* current stack depth */
        int trace;                  /* tied to $DB::trace (per-thread/proc) */
        int signal;                 /* tied to $DB::signal (per-thread/proc) */
                int baseoff;                /* offset from this struct to entire ring buffer base */
                watch_expr_t watches[4];    /* watch expressions */
                int cmdready;               /* 1 => command sent; -2 => response ready; 0 => empty */
                char command[4];            /* ext. command entry */
                int msglen;                 /* length of msg */
                char msgarea[];             /* ext. message area */
        ring_slot_t slots[];        /* slots */
    } ring_buffer_t;

When a thread exits, the DESTROY() method of the Devel::RingBuffer or Devel::RingBuffer::Ring object (whichever occurs first) will free the allocated ring buffer.

  • pid is the PID of the ring buffer owner.

  • tid is the thread ID of the ring buffer owner.

  • currSlot is the current "top" slot of the ring buffer, i.e., the slot which contains the name/linenumber/timestamp of the currently executing subroutine.

  • depth is the current depth of the stack; note that the depth may exceed the number of slots, which will cause the ring buffer to wrap and overwrite the oldest slots.

  • trace is tie'd to $DB::trace, thereby permitting per-thread single-stepping while some/all other threads/processes continue to run without interruption.

  • signal is tie'd to $DB::signal, to provide an additional signaling mechanism between the Monitor and individual threads of the AUT.

  • baseoff is the offset back to the base of the entire ring buffer. This field helps optimize accesses to global information that would otherwise require additional XS call parameters and SV translation. The XS methods use it internally to compute the address of the ring buffer header from an individual ring's address.

  • watches contains the watch list (see "Watch Expressions"). Currently, only 4 expressions per thread are supported.

  • cmdready, command, msglen, and msgarea provide the per-thread command/response message area between the AUT and Monitor, as described previously. When Monitor needs to post a command to an AUT thread, it waits until cmdready == 0, then posts the command by writing to the command, msglen, and msgarea fields, then setting cmdready = 1. AUT can either test, or wait for, a command by testing for cmdready == 1, then reading the command and message data and performing any needed command processing, then posting any response data, and setting cmdready == 0.

  • slots contains the individual subroutine slots (see next section).


        typedef struct {
                int linenumber;         /* current execution linenumber in subroutine */
                double timestamp;       /* Time::HiRes::time() timestamp of the execution */
                char subroutine[1];     /* name of subroutine (as reported by $DB::sub) */
                                /* size determine by the slot_sz parameter */
        } ring_slot_t;
  • linenumber is the script linenumber currently being executed within the slot's current subroutine. This field of the current slot is updated on each pass of the AUT through DB::DB().

  • timestamp is the Time::HiRes::time() value when the current linenumber of the current subroutine was executed. This field of the current slot is updated on each pass of the AUT through DB::DB().

  • subroutine is the name of subroutine (as reported by $DB::sub) assigned to the slot; it is set by DB::sub() when a new subroutine call is made.

Watch Expressions

Each ring includes a set of slots for use with watch expressions, defined as

        typedef struct {
        int inuse;          /* 1 => in use, -2 => freeing, 0 => freed */
        int exprlength;     /* length of expr text */
        char expr[256];     /* expr text */
        int resready;       /* 0 => monitor has read last result */
        int reslength;      /* result length before truncation;
                               length < 0 the expression eval failed, with
                               error text in the result area;
                               length == 0 means result was undef */
        char result[512];   /* result text */
        } watch_expr_t;

The Monitor can set an expression, and each AUT thread scans its watchlist for expressions to evaluate. A lazy concurrency control mechanism is used to minimize locking overhead:

  • Monitor locates the first entry with inuse == 0

  • Monitor writes out the epxression, and sets the expression length

  • Monitor sets inuse = 1

  • AUT scans its thread's watch slots looking for entries with (inuse == 1) and (resready == 0). When it finds an entry, it eval's the expression, writes the result to the slot (truncating if needed), and sets resready = 1. Note that the AUT won't eval the expression for every pass through DB::DB(), unless $DB::single is set to indicate single-stepping and the Monitor reads the watched expressions after every single-step.

  • Monitor scans the watchlist looking for entries with (inuse == 1) and (resready == 1). It reads and formats the result, then sets resready = 0 to indicate it is done with the entry.

  • To remove a watch item, Monitor sets inuse = -2

  • When AUT sees an entry with inuse == -2, it acknowledges the free operation by setting inuse = 0.


Refer to included classdocs for summary and detailed descriptions of methods.


  • support for 64 bit Perl/platforms

  • UTF 8 support/validation

  • More XS/C to further minimize its impact on the monitored application.






strace(1) (or truss(1))


Dean Arnold

Copyright(C) 2006, Dean Arnold, Presicient Corp., USA. All rights reserved.

Permission is granted to use this software under the same terms as Perl itself. Refer to the Perl Artistic License for details.