Issue diagnosed and reported by Ambroz Bizjak.
The dbContextReadNotifyCacheAllocator allocator clears
its cache of free buffers on an array size change.
But doesn't consider buffer already in use, which
will later be free'd. Such buffers were being
returned to the cache, then reused in allocations
for which they are too short.
Track the size of buffers which are in use.
Only return buffers with the present length
to the cache. Others are free'd immediately.
A dbCa link does a ca_get with type DBR_CTRL_DOUBLE
to populate its list of attribute values immediately
after connecting. If the target is a DBF_*LINK field
it used to return an error, preventing the link from
properly connecting. This change makes dbGetField()
return a single NAN value instead of rejecting the
request.
Fixes: lp:545358
Penalizing the 0.1 second scan thread by 0.1 seconds
guarantees another overrun in most cases...
Warning message improved.
Also added a release note for Jeff's CAS commit.
Reading back a .TSEL field that was originally pointed to
another record's .TIME field was exposing internal details
that didn't need to be made visible. This commit hides the
.TIME => .VAL change completely, and allows a .TSEL link
field to be saved and restored properly using autosave.
This is really a workaround for a Win32/MinGW bug in sscanf, which
will only set the %n argument for the format "%u %n" when there is
a space following the unsigned number.
This changes the IOC's behavior very slightly.
Clients try to use long string support to fetch DBF_STRING
fields use DBF_CTRL_CHAR with a 1-element array, but the IOC
was rejecting that. This permits it, and also ensures that
the resulting strings are zero-terminated.
Fixes lp:907761
o fixed launchpad epics base bug 878387
o added regression test for bug 878372
o build the optional test and diagnostic codes for all (not just vxWorks) IOCs
o added acctstRegister optional IOC shell registration for regression tests
This fixes a bug introduced with the dynamic-array branch.
The new version of db_get_field() was using the dbr_size_n() macro to
calculate the offset to the start of the buffer area to be zeroed, but
when nRequest is zero this gives the wrong result; the dbr_ structures
always contain at least one value element, and dbr_size_n() thus has
to include that in its size calculation.
The rsrv server is updated in this commit to deliver variable sized
waveforms in response to a zero length request, implementing an
extension to the v4 EPICS protocol tied to minor version 13.
The core of the changes are in read_reply, rsrv/camessage.c, which first
reserves packet space for the largest possible response, but then resizes
the packet according to how much data is actually returned from the
database.
