fix index when ndims is large

[wkliao]

Resolve the Segmentation fault. Below are the gdb traces.

Program terminated with signal SIGSEGV, Segmentation fault.
#0  0x00007f1a2aea2907 in ncmpi_put_vars_int_all (ncid=0, varid=0, start=0x19e08c0, count=0x198fbf0, stride=0x1991c00, 
    buf=0x7f1a25eec010) at ./darshan-pnetcdf-api.c:11151
11151	                    common_access_vals[1+i] = count[i];

#0  0x00007f2f784f393e in ncmpi_put_vars_int_all (ncid=0, varid=0, start=0x1c7b8c0, count=0x1c2abf0, stride=0x1c2cc00, 
    buf=0x7f2f7353d010) at ./darshan-pnetcdf-api.c:11153
11153	                        common_access_vals[1+i+PNETCDF_VAR_MAX_NDIMS] = stride[i];

#0  0x00007fa89f55307f in ncmpi_get_vars_int_all (ncid=0, varid=0, start=0xca38c0, count=0xc52bf0, stride=0xc54c00, 
    buf=0x7fa89a58b010) at ./darshan-pnetcdf-api.c:13080
13080	                    common_access_vals[1+i] = count[i];

[shanedsnyder]

Is this an example that's publicly available and that I could reproduce with? If not, do you know what exactly triggers it? I.e., any variable with more than 5 dimensions, or?

[wkliao]

The seg fault occurred when I built PnetCDF with option --enable-burst_buffering
which tests a variable with a large dimension size 1024 in test/burst_buffer/highdim.c.

[shanedsnyder]

Anything else special I need to trigger the seg fault? That test case runs fine for me now on Darshan main branch.

Are you running the test just via make check on your laptop? Or are you running the test directly with different parameters?

FYI, I'm using PnetCDF main branch for testing. I could try a specific version if that's what you're using, also.

[wkliao]

I ran 'make check' and 'make ptest' on a local Redhat server.
I suggest to always use an official release of PnetCDF, like in #874.

[shanedsnyder]

Confirmed on my end with the 1.12.3 release.

Proposed changes look right to me. I spot checked a few examples that use vara/vars interfaces and the affected counters look as expected, so all looks good here. Thanks!