|
NAME | C SYNOPSIS | DESCRIPTION | RUN-TIME CONTROL | EXAMPLES | ENVIRONMENT | SEE ALSO | DIAGNOSTICS | COLOPHON |
PMFAULT(3) Library Functions Manual PMFAULT(3)
__pmFaultInject, PM_FAULT_POINT, PM_FAULT_CHECK, __pmFaultSummary -
Fault Injection Infrastracture for QA
#include <pcp/pmapi.h>
#include <pcp/impl.h>
#include <pcp/fault.h>
void __pmFaultInject(const char *ident, int class);
void __pmFaultSummary(FILE *f);
PM_FAULT_POINT(ident, class);
PM_FAULT_CHECK(class);
cc -DPM_FAULT_INJECTION=1 ... -lpcp_fault
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):
PM_FAULT_ALLOC
Will cause the next call to malloc(3), realloc(3) or strdup(3)
to fail, returning NULL and setting errno to ENOMEM. We could
extend the coverage to all of the malloc-related routines, but
these three are sufficient to cover the vast majority of the
uses within libpcp.
PM_FAULT_PMAPI
Will cause the next call to a PMAPI routine to fail by
returning the (new) PCP error code PM_ERR_FAULT. At the this
stage, only __pmRegisterAnon(3) has been instrumented as a
proof of concept for this part of the facility.
PM_FAULT_TIMEOUT
Will cause the next call to an instrumented routine to return
the PCP error code PM_ERR_TIMEOUT. At this stage, only
__pmGetPDU(3) has been instrumented to check for this class of
fault injection.
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:
PM_FAULT_PMAPI
The enclosing routine will return immediately with the value
PM_ERR_FAULT - this assumes the enclosing routine is of type
int foo(...) like all of the PMAPI routines.
PM_FAULT_TIMEOUT
The enclosing routine will return immediately with the value
PM_ERR_TIMEOUT - this assumes the enclosing routine is of type
int foo(...) like all of the PMAPI routines.
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
__pmFaultInject 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 ops, 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 [0] flags 0x1 (point) ---
sample.event.param_string "fetch #0"
10:46:12.413 --- event record [1] 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 [0] 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
Refer to the PCP and PCP QA source code.
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.
PM_FAULT_CONTROL
Full path to the fault injection control file.
LD_PRELOAD
Force libpcp_fault to be used in preference to libpcp.
PMAPI(3)
Some non-recoverable errors are reported on stderr.
This page is part of the PCP (Performance Co-Pilot) project.
Information about the project can be found at ⟨http://www.pcp.io/⟩.
If you have a bug report for this manual page, send it to
pcp@oss.sgi.com. This page was obtained from the project's upstream
Git repository ⟨git://git.pcp.io/pcp⟩ on 2017-07-05. If you discover
any rendering problems in this HTML version of the page, or you
believe there is a better or more up-to-date source for the page, or
you have corrections or improvements to the information in this
COLOPHON (which is not part of the original manual page), send a mail
to man-pages@man7.org
Performance Co-Pilot PMFAULT(3)