* Revert "Remove Autotools sed hack (#3848)"
This reverts commit 8b3ffdef30.
* Fix libtool sed cleanup on MacOS
Convert sed -i line to sed > libtool.bak && mv libtool.bak libtool
to avoid non-portable -i option.
* Fix for github issue #1388: can't delete renamed dense attribute with corder tracking enabled
The problem occurs in step 3(b) below which will delete the attribute with corder x
from the creation order index v2 B-tree.
The rename sequence in H5A__dense_rename() occurs in the following order:
1) The old attribute with corder x was removed from the creation order index v2 B-tree
2) The new renamed attribute was inserted via H5A__dense_insert():
(a) insert the attribute with new name j into the name index v2 B-tree
(b) insert the attribute with corder x into the creation order index v2 B-tree
3) The old attribute was removed via H5A__dense_remove():
(a) remove the attribute with old name k from the name index v2 B-tree
(b) remove the attribute with coorder x from the creation order index v2 B-tree
Fix: deactivate the "corder_bt2_addr" field so that H5A__dense_remove()
won't delete the attribute with corder x from the creation order index v2 B-tree.
* Improve spec. reading superblock into cache (a little) by using v2 size
Instead of reading the absolute minimal possible, use the likely value of
a v2+ superblock w/8-byte addresses & lengths.
Document the limitation in the Passthrough Conncector section of the VOL Connector Author Guide.
The limitation is posted by Neil in the github issue on Dec 22, 2022.
Fixes a race condition where the reader opens the file and sets its EOF from the
file's size (from the stat() call in the driver open callback). Then, before
the reader can read the file's superblock, a SWMR writer races in, extends the
file, and closes the file, writing an updated superblock with the 'writer' and
'SWMR writer' flags in the superblock off (appropriately). Then the reader
proceeds to read the superblock, and flags the EOF as wrong. Taking out the
check for the 'writer' and 'SWMR writer' flags will cause SWMR readers to avoid
flagging the file as incorrect.
* Properly clean up cache when failing to load an object header
* Don't check message type a second time in H5G__open_oid if the first attempt returns error
* Add more asserts to H5O__assert() to avoid segfaults
---------
Co-authored-by: github-actions <41898282+github-actions[bot]@users.noreply.github.com>
All calls to the H5I routines are now made in API routines (sometimes in
FUNC_ENTER/LEAVE_* macros), except for some calls to H5E_clear_stack() within
the library, but I'm planning to remove those over time.
Also, made all the library internal error messages into static const variables,
instead of malloc'ing them, which means that they can just be referenced
and not copied.
Several new and updated auto-generated header files were necessary to enable
this.
* Fixes and cleanup for ph5diff
Fixes concurrency issues in ph5diff that can cause interleaved
output
Fixes an issue where output can sometimes be dropped if it ended
up in ph5diff's output overflow file
Fixes an issue where MPI_Init is called after HDF5 has been
initialized, preventing the library from setting up an MPI
attribute to perform cleanup on MPI_Finalize
Fixes an issue in config/cmake/runTest.cmake where the CMake
logic would try to access an invalid list index if the number
of lines in a test's output and reference files don't match
* Add release note
The buffer size for checksum was smaller than H5_SIZEOF_CHKSUM, causing an
overflow while calculating the offset to the checksum in the buffer.
A check was added so H5F_get_checksums would fail appropriately in all
of its occurrences.
Fix gh-4434
- Add Fortran H5R APIs:
h5rcreate_attr_f, h5rcreate_object_f, h5rcreate_region_f,
h5ropen_attr_f, h5ropen_object_f, h5ropen_region_f,
h5rget_file_name_f, h5rget_attr_name_f, h5rget_obj_name_f,
h5rcopy_f, h5requal_f, h5rdestroy_f, h5rget_type_f
- Fixed function H5Requal actually to compare the reference pointers
Fixed an issue with H5Requal always returning true because the
function was only comparing the ref2_ptr to itself.
* Fix for github issue #3790: infinite loop closing library
Cause of the problem:
When h5dump tries to open the user provided test file, the metadata cache will
call the "get_final_load_size" callback to find out the actual size of the
the root object header. The callback function will call
H5O__prefix_deserialize() to allocate space for the object header
data structure (via H5FL_CALLOC) and to deserialize the object header prefix
in order to find the actual size of the object header.
The metadata cache will then check whether the actual size obtained
will exceed the file's EOA.
Since the actual size obtained from the test file exceeds the EOA,
the metadata cache throws an error and return.
However, the oh structure that was allocated in H5O__prefix_deserialize()
was not freed and hence causing the problem described in this issue.
Fix:
1) Deallocate the oh structure after obtaining and saving the needed
information in udata which will be used later on in the "verify_chksum" callback.
2) Deserialize the object header prefix in the "object header's
"deserialize" callback regardless. The original coding intends to keep the
deserialized prefix so that the object header's "deserialize" callback
does not need to deserialize the prefix again if the object header is coming
through the "get_final_load_size" callback.
