void __pmFaultInject(const char *
void __pmFaultSummary(FILE * f );
PM_FAULT_CHECK( class );
As part of the coverage-driven changes to QA in PCP 3.6, it became apparent that we needed someway to exercise the ``uncommon'' code paths associated with error detection and recovery.
The facilities described below provide a basic fault injection infrastructure (for libpcp only at this stage, alhough the mechanism is far more general and could easily be extended).
A special build is required to create libpcp_fault and the associated < pcp/fault.h > header file. Once this has been done, new QA applications may be built with -DPM_FAULT_INJECTION=1 and/or existing applications can be exercised in presence of fault injection by forcing libpcp_fault to be used in preference to libpcp as described below.
In the code to be tested, __pmFaultInject defines a fault point at which a fault of type class may be injected. ident is a string to uniquely identify the fault point across all of the PCP source code, so something like "libpcp/" __FILE__ ":<number>" works just fine. The ident string also determines if a fault will be injected at run-time or not - refer to the RUN-TIME CONTROL section below. class selects a failure type, using one of the following defined values (this list may well grow over time):
To allow fault injection to co-exist within the production source code, PM_FAULT_POINT is a macro that emits no code by default, but when PM_FAULT_INJECTION is defined this becomes a call to __pmFaultInject . Throughout libpcp we use PM_FAULT_POINT and not __pmFaultInject so that both libpcp and libpcp_fault can be built from the same source code.
Similarly, the macro PM_FAULT_CHECK emits no code unless PM_FAULT_INJECTION is defined, in which case if a fault of type class has been armed with __pmFaultInject then, the enclosing routine will trigger the associated error behaviour. For the moment, this only works for the following class types:
A summary of fault points seen and faults injected is produced on stdio stream f by __pmFaultSummary .
Additional tracing (via -Dfault and DBG_TRACE_FAULT ) and a new PMAPI error code ( PM_ERR_FAULT ) are also defined, although these will only ever be seen or used in libpcp_fault . If DBG_TRACE_FAULT is set the first time __pmFaultInject is called, then __pmFaultSummary will be called automatically to report on stderr when the application exits (via atexit (3)).
Fault injection cannot be nested. Each call to __pmFaultInject clears any previous fault injection that has been armed, but not yet executed.
The fault injection infrastructure is not thread-safe and should only be used with applications that are known to be single-threaded.
By default, no fault injection is enabled at run-time, even when __pmFaultInject is called.
Faults are selectively enabled using a control file, identified by the environment variable $PM_FAULT_CONTROL ; if this is not set, no faults are enabled.
The control file (if it exists) is read the first time
is called, and
contains lines of the form:
ident op number
that define fault injection guards.
ident is a fault point string (as defined by a call to __pmFaultInject , or more usually the PM_FAULT_POINT macro). So one needs access to the libpcp source code to determine the available ident strings and their semantics.
op is one of the C-style operators >= , > , == , < , <= , != or % and number is an unsigned integer. op number is optional and the default is >0
The semantics of the fault injection guards are that each time __pmFaultInject is called for a particular ident , a trip count is incremented (the first trip is 1); if the C-style expression tripcount op number has the value 1 (so true for most op s, or the remainder equals 1 for the % op ), then a fault of the class defined for the fault point associated with ident will be armed, and executed as soon as possible.
Within the control file, blank lines are ignored and lines beginning with # are treated as comments.
For an existing application linked with libpcp fault injection may still be used by forcing libpcp_fault to be used in the place of libpcp . The following example shows how this might be done.
$ export PM_FAULT_CONTROL=/tmp/control $ cat $PM_FAULT_CONTROL # ok for 2 trips, then inject errors libpcp/events.c:1 >2 $ export LD_PRELOAD=/usr/lib/libpcp_fault.so $ pmevent -Dfault -s 3 sample.event.records host: localhost samples: 3 interval: 1.00 sec sample.event.records[fungus]: 0 event records __pmFaultInject(libpcp/events.c:1) ntrip=1 SKIP sample.event.records[bogus]: 2 event records 10:46:12.413 --- event record  flags 0x1 (point) --- sample.event.param_string "fetch #0" 10:46:12.413 --- event record  flags 0x1 (point) --- sample.event.param_string "bingo!" __pmFaultInject(libpcp/events.c:1) ntrip=2 SKIP sample.event.records[fungus]: 1 event records 10:46:03.416 --- event record  flags 0x1 (point) --- __pmFaultInject(libpcp/events.c:1) ntrip=3 INJECT sample.event.records[bogus]: pmUnpackEventRecords: Cannot allocate memory __pmFaultInject(libpcp/events.c:1) ntrip=4 INJECT sample.event.records[fungus]: pmUnpackEventRecords: Cannot allocate memory __pmFaultInject(libpcp/events.c:1) ntrip=5 INJECT sample.event.records[bogus]: pmUnpackEventRecords: Cannot allocate memory === Fault Injection Summary Report === libpcp/events.c:1: guard trip>2, 5 trips, 3 faults
src/libpcp/src/derive.c uses PM_FAULT_CHECK .
src/libpcp/src/err.c and src/libpcp/src/events.c use PM_FAULT_POINT .
src/libpcp/src/fault.c contains all of the the underlying implementation.
src/libpcp_fault contains the recipe and Makefile for creating and installing libpcp_fault and < pcp/fault.h > .
QA/477 and QA/478 show examples of control file use.
Some non-recoverable errors are reported on stderr .