Segmentation fault when reading an array of fixed length strings (inc in ncdump)

[IanHolthamGMV]

I have been writing a file handler in C++ for HDF5 files with v4.8.1 on Linux (Centos7).

Everything has gone well until I tried to read an array of fixed length strings.

I created a test file with a single Group and then a single Dataset of strings of fixed length 20 in a 1 dimensional array of 5 strings. If I then call ncdump on that file I get a segmentation fault (output from ncdump at the end of this post). Stepping through my own code with the same file reveals a segmentation fault when I call nc_free_string on this dataset.

I've stepped into the code and looking at line 1875 of hdf5var.c, it seems to allocate memory for just the first string in the array (fstring_len is 20):

if (!(*(char **)data = malloc(1 + fstring_len)))

Is there something I've misunderstood, or is this a bug in the library?

Thanks

Ian

StringTest.txt

[DennisHeimbigner]

I think I see the problem. Netcdf-4 does not support HDF5 style fixed length strings; it only
supports variable length strings (more or less so when you do this:

dimensions: phony_dim_0 = 5 ;
variables: string StringArray(phony_dim_0) ;

You are in effect doing the equivalent of allocated a 5 element array each of wholes elements
is a variable length string. More or less equivalent to "char* StringArray[5]"
Unfortunately, F90 does not handle this case well. You might be better off doing something
like this:

dimensions: strlen = 20; nstrings=5;
variable: char StringArray(nstrings,strlen)

[jswhit]

4.9.0 segfaults when trying to read an hdf5 file with fixed length string arrays (test file available at Unidata/netcdf4-python#1172).

[DennisHeimbigner]

I have a fix that I am testing now. Your identification of

if (!(*(char **)data = malloc(1 + fstring_len)))

as the source of the problem appears to be correct, although the fix
is a bit complicated.