cdata finalizer called twice
- From: Peter Colberg <peter@xxxxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Wed, 12 Nov 2014 23:14:29 -0500
Hi Mike,
I am tracing a possible bug in the LuaJIT 2.0 branch where a cdata
finalizer is called twice for the same object. The cdata type is an
incomplete struct defined as
ffi.cdef[[
typedef struct _cl_event *cl_event;
]]
and the cdata objects are constructed using a function similar to
ffi.cdef[[
cl_int clEnqueueMarker(cl_command_queue, cl_event *);
cl_int clReleaseEvent(cl_event);
]]
local function release_event(event)
local status = C.clReleaseEvent(event)
if status ~= C.CL_SUCCESS then
io.write("DUPLICATE clReleaseEvent("..tostring(event)..")\n")
return error(errors[status]) -- Invalid event
end
io.write("clReleaseEvent("..tostring(event)..")\n")
end
local cl_event_1 = ffi.typeof("cl_event[1]")
local function enqueue_marker(queue)
local event = cl_event_1()
local status = C.clEnqueueMarker(queue, event)
if status ~= C.CL_SUCCESS then return error(errors[status]) end
return ffi.gc(event[0], release_event)
end
When I run »10000 iterations of creating and destroying cl_event
objects, forcing immediate destruction through collectgarbage(),
LuaJIT will for most runs (but not always) eventually call the
finalizer of a few of these objects twice:
…
clReleaseEvent(cdata<struct _cl_event *>: 0x038fd658)
clReleaseEvent(cdata<struct _cl_event *>: 0x037bb198)
clReleaseEvent(cdata<struct _cl_event *>: 0x038f9b28)
clReleaseEvent(cdata<struct _cl_event *>: 0x03ba0a88)
clReleaseEvent(cdata<struct _cl_event *>: 0x00960b48)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bcad48)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bca4b8)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bca168) <---
DUPLICATE clReleaseEvent(cdata<struct _cl_event *>: 0x03bca168) <--!
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc9e08) <---
DUPLICATE clReleaseEvent(cdata<struct _cl_event *>: 0x03bc9e08) <--!
/theory2/u4/pcolberg/vcs/git/luajit-2.0/src/luajit: Invalid event
clReleaseEvent(cdata<struct _cl_event *>: 0x03ba0448)
clReleaseEvent(cdata<struct _cl_event *>: 0x03ba1268)
clReleaseEvent(cdata<struct _cl_event *>: 0x038fd5e8)
clReleaseEvent(cdata<struct _cl_event *>: 0x038fb678)
clReleaseEvent(cdata<struct _cl_event *>: 0x00960b48)
clReleaseEvent(cdata<struct _cl_event *>: 0x038fb888)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc8998)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc8068)
clReleaseEvent(cdata<struct _cl_event *>: 0x039e6678)
clReleaseEvent(cdata<struct _cl_event *>: 0x038f9b28)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc8ed8)
clReleaseEvent(cdata<struct _cl_event *>: 0x03814988)
clReleaseEvent(cdata<struct _cl_event *>: 0x038fbdb8)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc9748)
clReleaseEvent(cdata<struct _cl_event *>: 0x038f2068)
clReleaseEvent(cdata<struct _cl_event *>: 0x03bc9aa8)
clReleaseEvent(cdata<struct _cl_event *>: 0x03ba0658)
clReleaseEvent(cdata<struct _cl_event *>: 0x039e6b08)
Segmentation fault (core dumped)
Using git bisect I identified this commit as causing the issue:
commit 0b55e05d06bc760070b8ca9f99c01c3d8e67e7db
Author: Mike Pall <mike>
Date: Wed Oct 10 18:56:16 2012 +0200
FFI: Compile ffi.gc().
Hopefully the above information provides sufficient hints, since so
far I could not produce a self-contained minimal example.
Thanks,
Peter
Other related posts:
