BUG: Fixing memory leaks in read_csv

* Move allocation of na_hashset down to avoid a leak on continue
* Delete na_hashset if there is an exception
* Clean up table before raising an exception

Closes gh-21353.

doc/source/whatsnew/v0.24.0.rst

@@ -1382,6 +1382,7 @@ Notice how we now instead output ``np.nan`` itself instead of a stringified form
 - Bug in :func:`DataFrame.to_string()` that caused representations of :class:`DataFrame` to not take up the whole window (:issue:`22984`)
 - Bug in :func:`DataFrame.to_csv` where a single level MultiIndex incorrectly wrote a tuple. Now just the value of the index is written (:issue:`19589`).
 - Bug in :meth:`HDFStore.append` when appending a :class:`DataFrame` with an empty string column and ``min_itemsize`` < 8 (:issue:`12242`)
+- Bug in :func:`read_csv()` in which memory leaks occurred in the C engine when parsing ``NaN`` values due to insufficient cleanup on completion or error (:issue:`21353`)
 - Bug in :func:`read_csv()` in which incorrect error messages were being raised when ``skipfooter`` was passed in along with ``nrows``, ``iterator``, or ``chunksize`` (:issue:`23711`)
 - Bug in :meth:`read_csv()` in which :class:`MultiIndex` index names were being improperly handled in the cases when they were not provided (:issue:`23484`)
 - Bug in :meth:`read_html()` in which the error message was not displaying the valid flavors when an invalid one was provided (:issue:`23549`)

pandas/_libs/parsers.pyx

@@ -1070,18 +1070,6 @@ cdef class TextReader:
 
             conv = self._get_converter(i, name)
 
-            # XXX
-            na_flist = set()
-            if self.na_filter:
-                na_list, na_flist = self._get_na_list(i, name)
-                if na_list is None:
-                    na_filter = 0
-                else:
-                    na_filter = 1
-                    na_hashset = kset_from_list(na_list)
-            else:
-                na_filter = 0
-
             col_dtype = None
             if self.dtype is not None:
                 if isinstance(self.dtype, dict):
@@ -1106,13 +1094,26 @@ cdef class TextReader:
                                               self.c_encoding)
                 continue
 
-            # Should return as the desired dtype (inferred or specified)
-            col_res, na_count = self._convert_tokens(
-                i, start, end, name, na_filter, na_hashset,
-                na_flist, col_dtype)
+            # XXX
+            na_flist = set()
+            if self.na_filter:
+                na_list, na_flist = self._get_na_list(i, name)
+                if na_list is None:
+                    na_filter = 0
+                else:
+                    na_filter = 1
+                    na_hashset = kset_from_list(na_list)
+            else:
+                na_filter = 0
 
-            if na_filter:
-                self._free_na_set(na_hashset)
+            try:
+                # Should return as the desired dtype (inferred or specified)
+                col_res, na_count = self._convert_tokens(
+                    i, start, end, name, na_filter, na_hashset,
+                    na_flist, col_dtype)
+            finally:
+                if na_filter:
+                    self._free_na_set(na_hashset)
 
             if upcast_na and na_count > 0:
                 col_res = _maybe_upcast(col_res)
@@ -2059,6 +2060,7 @@ cdef kh_str_t* kset_from_list(list values) except NULL:
 
         # None creeps in sometimes, which isn't possible here
         if not isinstance(val, bytes):
+            kh_destroy_str(table)
             raise ValueError('Must be all encoded bytes')
 
         k = kh_put_str(table, PyBytes_AsString(val), &ret)