Broken isinf on Solaris 9

[human]

The current implementation of isinf on Solaris 9 leads to the problem that valid line numbers are rejected as "out of range" as outlined in this bug reported for Orca:

•  https://www.opencsw.org/mantis/view.php?id=4380

The error message is

  LINE width '1:average1#0000ff:Zero-fill-on-demand' is out of range in line 'LINE1:average1#0000ff:Zero-fill-on-demand'

The code leading to this message is from rrd_graph_helper.c:

    case GF_LINE:
        if (c1 == ':') {
            gdp->linewidth = 1;
            dprintf("- using default width of 1\n");
        } else {
            i = 0;
            sscanf(&line[*eaten], "%lf:%n", &gdp->linewidth, &i);
fprintf(stderr,"LF: %lf\n", gdp->linewidth);
fprintf(stderr,"I: %ld %ld %ld\n",fpclass(gdp->linewidth),FP_SNAN,FP_QNAN);
            if (!i) {
                rrd_set_error("Cannot parse line width '%s' in line '%s'\n",
                              &line[*eaten], line);
                return 1;
            } else {
                dprintf("- scanned width %f\n", gdp->linewidth);
                if (isnan(gdp->linewidth)) {
                    rrd_set_error
                        ("LINE width '%s' is not a number in line '%s'\n",
                         &line[*eaten], line);
                    return 1;
                }
                if (isinf(gdp->linewidth)) {
                    rrd_set_error
                        ("LINE width '%s' is out of range in line '%s'\n",
                         &line[*eaten], line);
                    return 1;
                }
                if (gdp->linewidth < 0) {
                    rrd_set_error
...

The problem is the broken implementation of 'isinf' for Solaris 9 in rrd_config_bottom.h:

/* for Solaris */
#if (! defined(HAVE_ISINF) && defined(HAVE_FPCLASS))
#  define HAVE_ISINF 1
#  ifdef isinf
#  undef isinf
#  endif
#  define isinf(a) (!!(fpclass(a) & (FP_SNAN|FP_QNAN)))
#endif

The problem is that fpclass(a) does not return a bitmask, but a number indicating the class of the floatingpoint number as documented in fpclass(3c):

     The fpclass() function returns one of the following  classes
     to which dsrc belongs:

These are from /usr/include/ieeefp.h and valid results:

typedef enum    fpclass_t {
        FP_SNAN = 0,    /* signaling NaN */
        FP_QNAN = 1,    /* quiet NaN */
        FP_NINF = 2,    /* negative infinity */
        FP_PINF = 3,    /* positive infinity */
        FP_NDENORM = 4, /* negative denormalized non-zero */
        FP_PDENORM = 5, /* positive denormalized non-zero */
        FP_NZERO = 6,   /* -0.0 */
        FP_PZERO = 7,   /* +0.0 */
        FP_NNORM = 8,   /* negative normalized non-zero */
        FP_PNORM = 9    /* positive normalized non-zero */
} fpclass_t;

This alternative implementation for isinf solves the problem:

#define isinf(a) (fpclass(a) == FP_NINF || fpclass(a) == FP_PINF)

[oetiker]

indeed ... bad bug ... a regression introduced by a bad optimization added r1904 ... fixed in r2093 ... thanks