GT.M Version 5.2-000B Release Notes

In UTF8 mode, functionality related to Unicode™ becomes available and standard string-oriented operations (particularly intrinsic functions) operate with UTF-8 encoding.

In the default M mode, GT.M exhibits the same behavior as previous versions.

In response to two incidents of repeated GVxxxFAIL or TPFAIL errors with failure code xLLL caused by process, rather than database, damage, the condition that led to the xLLL failure now triggers a GTMASSERT. This allows FIS to get better diagnostic information and prevents the damaged process from failing repeatedly. (S9H02-002641)

GT.M now issues a VERMISMATCH error whenever a process tries to access a GT.M database that is already being accessed by a different version of GT.M. This is both in UNIX and OpenVMS. Previous versions of GT.M used to issue a different message that was less user-friendly. In addition prior versions of GT.M in OpenVMS could potentially cause deadlocks and even database damage if GT.M versions were mixed. All these issues are now fixed. (C9D08-002387)

GT.M manages shared resources through the use of shared memory. For performance reasons, a significant part of this management is based on algorithms designed to minimize or avoid locks. As processors, caches, and multi-processor interconnects have evolved, computer architectures have become more prone to memory update patterns that appear to a reading process in a different order than they do in the code of the updating process (commonly called out-of-order stores). GT.M uses platform specific primitives (commonly called memory barriers) to prevent out-of-order store issues where they pose problems. GT.M adopted memory barriers some time ago in selected places within the database logic. We recently found that the journal code, which is subject to less intense inter-process contention, needs some as well. This release adds a few memory barriers to the journal management logic to prevent rare problems, including journal file corruption, which could be induced by out-of-order stores. The likelihood of encountering this issue is function of both the particular hardware (different models of the same CPU line may have different characteristics) and the level of contending activity. [AIX, HP-UX, Tru64, OpenVMS] (C9E02-002511)

GT.M now prevents a possible source of database damage by ensuring that a non-programmatic process termination, typically due to MUPIP STOP, does not try to commit a transaction where the process has initialized in preparation for a transaction but has not yet formed any real updates. (C9F09-002757)

When GT.M and its utilities (source server etc.) read the replication instance file and find that the instance file was created in an older format, they issue a REPLINSTFMT error. Prior versions of GT.M used to leave an ipc semaphore (the ftok semaphore of the replication instance file) hanging around. This is now fixed so no semaphores are left around after issuing the REPLINSTFMT error. [UNIX] (C9G06-002800)

TPRESTART logging to the operator log used to be controlled by setting the logicals GTM_TPRESTART_LOG_START and GTM_TPRESTART_LOG_DELTA appropriately. This stopped working in V50000D onwards and used to cause any restarting process to ACCVIO if the above logicals were defined. This is now fixed to work as it used to. [OpenVMS] (C9H03-002829)

GT.M now protects against inappropriately long resource waits which could cause brief moments of degraded performance. [OpenVMS] (C9H03-002830)

GT.M no longer issues a GTMASSERT in the very rare case that a process uses the DAL call to disconnect all databases and then terminates due to a forced exit or an error before successfully [re-]opening any database.[OpenVMS] (C9H03-002833)

SET of global variables within a parenthesized list of targets where the global references use extended syntax containing one or more local variables now occurs properly when the local variables are defined or gives an appropriate UNDEF error when a local variable is undefined. In previous versions, even when the local variables were all defined, it used to give a message that did not specify the variable name (HPUX/AIX) or a segmentation fault (SIG-11). [Linux/Tru64/Solaris/AlphaVMS] (S9C08-002197)

Output-only parameters of type char * and string * without pre-allocation now trigger the ZCNOPREALLOUTPAR error. Also see C9F09-002753.[UNIX] (S9C10-002243)

Note that this change makes V5.2-000A not backward compatible for programs that use this feature. Although GT.M is normally fastidious with respect to maintaining backward compatibility, improvements in performance and robustness were judged to warrant an exception in this case.

The pattern match operator no longer causes a process to hang if the pattern following the question-mark (?) is a string literal more than 6 characters long and the repetition count consists of a period (.) with no leading and trailing integer literals. (S9F01-002526)

Unrecognized Intrinsic Functions and Intrinsic Special Variables and unrecognized deviceparameter constructions that are bypassed due to FALSE postconditionals no longer cause run-time errors. The error messages affected by this change are INVFCN, INVSVN, FNOTONSYS, SVNOSET, SVNONEW, DEVPARINAP, DEVPARUNK, DEVPARVALREQ. These errors still trigger error reports at compile time—for example "SET:0 $ZZZZ=1" continues to generate a compilation error— but no longer generates a run-time error. (S9F10-002572)

Extensive testing validates that two argument $STACK() and ZSHOW “S” work properly. If $ECODE is null then the two argument $STACK() returns the same result as ZSHOW "S". The result differs only when $ECODE contains error information. Fidelity recommends applications reset $ECODE in their error handlers before restoring the normal flow of control. (S9F12-002579)

READ X:0 now leaves $TEST set in case the only character READ is a terminator. (S9F12-002582)

Certain cases of an untimed OPEN of the TCP device that would cause the process to spin in a tight loop no longer do so.(S9G09-002625)

READ *X:<timeout>now properly times out when the timeout is greater than 8 seconds. [Linux] (S9G11-002628)

$zgetjpi(“”,”INVALID ARG”) now produces the BADJPIPARAM error rather than a segmentation fault (SIG-11). (S9G12-002632) [UNIX]

Pre-allocation of xc_string_t is now allowed and is required for output-only parameters on external calls. Also see S9C10-002243. (S9G12-002634, C9F09-002753) [UNIX]

ZSHOW "V" or ZWR when variables with "long" content no longer give a segmentation fault (SIG-11). This was fixed in V5.2-000 but the fix was not documented. (S9G12-002635)

$ZB now contains the correct value after a READ *. In particular, $ZB holds the full escape sequence after entry of an escape sequence when ESCAPE processing enabled. This was broken in GT.M V5.2-000. [AIX/Solaris/HPUX] (S9H02-002642)

Exponents greater than 18 digits no longer result in a numeric evaluation of 0; for example in previous versions, WRITE +"1E12234567890123456789" produced 0. (S9F11-002573)

The GT.M utilities and the mumps commands now use a space-dash (“ –“)combination to delimit/introduce a qualifier, which is standard UNIX usage, where they used to require only a dash ("-", which was not standard).

	This will cause existing shell scripts that use the idiosyncratic dash-only form to break – they can be fixed by adding one or more spaces in front of any dash.

This change allows the use of dashes in argument and qualifier values. [UNIX] (C9B04-001681)

$ORDER() now returns the correct value even if the optional second argument is specified and contains a global reference. In such cases, previous versions used to perform the $ORDER() operation on the global reference in the second argument instead of the first.

$REFERENCE is now set properly after a $ORDER() operation involving a global reference in the first or second argument. Previous versions used to set this incorrectly. For example $ORDER(^x(1)) used to incorrectly set $REFERENCE to ^x(1.00) instead of ^x(1). (C9B10-001744)

$ORDER() now returns the correct value where the second argument is an extrinsic function that manipulates the first argument. In such cases, previous versions could give an incorrect result. For example, $ORDER(V(2,-1),$$ABC) where the extrinsic $$ABC() sets the node V(2,2) to 1 and quits with a value of 1 now returns a result of 2 where it used to incorrectly return a null string (""). Substituting ^V for V in the example demonstrates the change for global references. (C9B10-001765)

$GET() with two arguments where both the arguments contain global references now sets the naked indicator correctly to the last global reference in the second argument. In cases where the first argument had a value, previous versions used to incorrectly set the naked indicator to that first argument. (C9B10-001767)

The processing of SET commands with multiple targets and $Piece() / $Extract() now complies better with the M standard.

GT.M now allows the $PIECE() and $EXTRACT functions as targets in a SET command that has a parenthesized, comma-separated list of targets. Previous versions disallowed this usage.

When SET arguments have multiple parenthesized (set-left) targets and a target is used as a subscript in more than one item in the list of targets that follow, ALL the targets now use the before-SET value (not the after-SET value) in conformance to the M-standard. Previous versions of GT.M used the before-SET value only for the last such subscripted target item in the list and incorrectly used the after-SET value for all prior subscripted target items. For example, set a=0,(a,array1(a),array2(a))=2 now sets both array1(0) and array2(0) to 2 whereas previous versions of GT.M would set array2(0) to 2 and array1(2) (instead of array1(0)) to 2.

If a SET target is of the form $PIECE(glvn,d,m,n) or $EXTRACT(glvn,m,n) GT.M now evaluates the right hand side of the Set even when m > n or n < 1. Previous versions of GT.M would not evaluate the right-hand-side of the SET which could result in the naked indicator not being set correctly. For example, set ^x=1,$piece(^a,";",3,2)=^b now correctly sets the naked indicator to point to the global ^b. Previous versions of GT.M set the naked indicator to ^x in this example. (C9C05-002003)

GT.M socket devices now allow up to the number of clients specified in the environment variable gtm_max_sockets. The default limit is 64, and there is an absolute maximum which at this point in human history exceeds your wildest dreams. $VIEW("MAX_SOCKETS") returns the current value. (C9C10-002149)

ZWRITE with a range (n:m) no longer no longer occasionally fails with a segmentation fault (SIG-11). [AIX; Solaris] (C9D01-002239)

The maximum length for the argument to the ZSYSTEM command has increased to 4K bytes.[UNIX] (C9E09-002637)

Messages and Recovery Procedures Manual is now current up to V5.2-000A. See http://www.fidelityinfoservices.com/user_documentation/V52AMsgRecProc/index.htm. (C9F07-002745)

OPEN of TCP device with an invalid address specification now gives an appropriate error rather than terminating the process. (C9G12-002808)

Pattern match with multiple unbounded patterns where a lower bound after the first unbounded pattern is 7 (9 for V5.2-000) or more now gives the correct result. For example, in GT.M V5.2 (1.1?.n1"."9.n), and, in prior versions, (1.1?.n1"."7.n) incorrectly returned 1. (C9G12-002813)

GT.M now reports complete error messages where in V5.2-000 utf-8 mode they were terminated at the first non-ASCII character. [HP-UX] (C9G12-002815)

GT.M no longer inappropriately issues a GTM-E-GTMSECSHRSOCKET when it finds a socket conflict (usually because of an abandoned socket) for use in activating GTMSECSHR actions.[UNIX] (C9H03-002828)

GT.M introduces a new deviceparameter called MOREREADTIME that applies to OPEN and USE commands for a SOCKET device. Except when a READ has a zero (0) timeout, MOREREADTIME specifies the time (in milliseconds) that a READ waits for input.

	This change may affect the operational behavior of TCP/IP aspects of your application. If the impact is adverse, specify a value of 200 milliseconds to restore the previous behavior.

TCP/IP is a stream-based protocol that guarantees that bytes arrive in the order in which they are transmitted, but does not make any further guarantees that they will be grouped in the same packets, etc. Usually, a higher-level communications / messaging protocol layered on top of TCP implements messages using one of the following approaches:

To implement approach 1 in MUMPS, specify a delimiter in an OPEN or USE command for the SOCKET device and then use a generic READ.

To implement approach 2, use a pair of fixed-length READs (READ #). Specifying a large value of MOREREADTIME is appropriate for Approach 1 and 2, but it tends to make Approach 3 work sub-optimally (except when it is implemented with zero timeout READ).

Specify a short value of MOREREADTIME for Approach 3. However, when used together with Approach 1 and 2, it uses CPU less efficiently than otherwise.

	Never set MOREREADTIME to 0 as it may cause a CPU to "spin". (C9H03-002835)

READ * in M-mode on Sequential Disk file and Socket device of a character with encoding between 128 to 255 inclusive now returns the correct code value. This fixes a problem introduced in V5.2-000 that resulted in READ * on those devices returning an incorrect (negative) value. (C9H04-002843)

The CHECKHEALTH, SHOWBACKLOG and SHUTDOWN qualifiers in a MUPIP REPLIC -SOURCE command now honor the value (if any) specified in the environment variable gtm_repl_instsecondary. [UNIX] (S9G07-002618)

MUPIP RUNDOWN now correctly uses an INFO message severity at its termination which prevents its hanging. This problem was introduced in V5.1-000 and is now fixed. [OpenVMS] (S9G07-002619)

MUPIP ENDIANCVT no longer terminates with an error while trying to convert a V5.0 format database. (S9G08-002621)

MUPIP INTRPT command no longer causes the loss of data in the input buffer on a terminal READ or direct mode.

If an interrupt (sent by the MUPIP INTRPT command) triggers code that does not resume from where the process was interrupted while a terminal READ is in progress, GT.M discards all input that the terminal READ received up to the time of that interrupt. ZSHOW "D" in a $ZINTERRUPT routine now includes a "ZINTERRUPT" designation for a terminal device if MUPIP INTRPT interrupted input on a terminal READ. Fidelity recommends that applications declare $ZINTERRPT routine specific $ETRAP/$ZTRAP error handlers in the interrupt processing code.

MUPIP INTRPT no longer causes the loss of previously received input on a terminal READ or direct mode. While GT.M processes an interrupt through the MUPIP INTRPT command, the only operation that GT.M permits on the current device ($IO) is the USE command with no deviceparameters. All other actions (CLOSE, OPEN, READ, WRITE) to that device trigger the ZINTRECURSEIO error. GT.M also triggers the ZINTDIRECT error for all attempts to BREAK into direct mode while that interrupt is processed. This allows the interrupt processing code to use another device and then restore the original device in an undisturbed state so that the interrupted application can resume as if uninterrupted (at least as far as the device behavior is concerned).[UNIX] (S9G12-002636)

Both socket /WAIT (server side) and socket READ now respond to MUPIP INTRPT. A socket device performing a client connect operation does not currently respond to an INTRPT until the connection is complete. (C9G04-002779, S9G03-002602)

	Note that the TCP device is not changed and still does not respond to INTRPT. FIS recommends replacing the deprecated TCP device usage with the standard SOCKET device.

MUPIP ENDIANCVT now ensures that buffers in shared memory are flushed to disk and completes close of its output file before printing success. (C9G06-002799)

If the environment variable gtm_quiet_halt is set to 1, then if the process is stopped via MUPIP STOP or by a SIGTERM signal (as is sent by some web servers), the FORCEDHALT message will be suppressed as it is assumed that this is the "normal" way this type of process will terminate. Full robust process termination, database flush, cleanup etc will be performed as is normally done. [UNIX] (C9G12-002812)

A Journal file's name can now include characters in Unicode™.[HP-UX] (C9G12-002814)

In rare cases where processes accessing a GT.M database were terminated abnormally, MUPIP RUNDOWN on that database could terminate prematurely with a SIG-11. This is now fixed.[UNIX] (C9F09-002755)

Users now have the option to install GT.M with or without Unicode-related functionality. GT.M now installs components related to the M-mode only on a computer which does not have ICU. GT.M continues to require ICU for functionality related to UTF-8 mode. (C9H02-002821)

GT.M now produces correct process initiation (“pini”) records in a journal file created during a recovery from an event that happens to interrupt a database file extension, where in prior versions this rare occurrence could result in an incorrect record. Such an incorrect record would not generally be noticed in M mode, but could cause a BADCHAR error in UTF-8 mode. (C9H03-002832)

In a backward rollback/recovery GT.M now avoids counting blocks requiring a V4 to V5 upgrade more than once when the action is interrupted and reissued more than once. In prior versions, a MUPIP INTEG of the recovered database could report a benign DBBTUWRNG integrity error, which it then automatically fixed. [UNIX] (C9H03-002836)

The INSTALL script no longer terminates prematurely on some HP-UX systems.[HP-UX] (S9G11-002627)

DSE ALL -DUMP [-ALL] now dumps the fileheader for all regions in the global directory. The -ALL qualifier selects the comprehensive format. The default is the briefer format. The -DUMP qualifier can only be intermixed with the -ALL qualifier. (C9G04-002784)

DBCERTIFY CERTIFY now correctly handles all conditions that lead to a split in the root level of a global variable tree. Previously these (fortunately rare) conditions triggered occasional spurious DBCMODBLK2BIG errors and the only workaround was to revert to a V4 format database and then repeat the scan/upgrade process. DBCERTIFY now always operates in M mode, while in V5.2-000 it could inappropriately attempt to operate in UTF-8 mode.

Also, the V5CBSU utility no longer triggers an occasional spurious INVFCN errors when it encounters a character with encoding in the range of 128-255. This problem was specific to V5.2-000 version only. [UNIX] (C9G04-002790)