K
Kirk Haines
Nutshell: The 0.8.5 version of the C extension for FCGI for Ruby leaks,
badly. I have a patch file at http://enigo.com/downloads/fcgi.c.patch if
you want to skip the explanation below and go right to it.
-----
http://enigo.com/projects/iowa/fcgipatch.html
When using the 0.8.5 version of the Ruby FCGI module with the C extension
instead of the pure Ruby extension, I noticed a substantial memory leak, to
the tune of around 16k+/request. A little poking through the code turned up
two seperate regions where memory is allocated without any provision to
free it later.
The first such instance occurs in the
fcgi_s_accept()
function. Memory is allocated to hold the FCGI request structure
(FCGX_Request), with this line:
req = ALLOC(FCGX_Request);
Then a data struct (fcgi_data), which has slots to hold the FCGI request,
the data streams, and the environment variables, is turned into a Ruby
object with a call to
Data_Make_Struct
A mark function is provided that marks the stream and environment slots for
the garbage collector. The request slot, however, is not marked since it
points to a C struct and not to a Ruby object. However, a free function is
not provided to cleanup either the root object or the request structure
that it contains. This is the first memory leak. It is fixed by providing a
function that will free this memory. This bug, with the same fix, was also
mentioned on ruby-core
(http://www.ruby-talk.org/cgi-bin/scat.rb/ruby/ruby-core/3245)last year in
reference to version 0.8.4.
Add this:
static void fcgi_free_req(fcgi_data *data)
{
free(data-<req);
free(data);
}
And then replace this:
obj = Data_Make_Struct(self, fcgi_data, fcgi_mark, 0, data);
with this:
obj = Data_Make_Struct(self, fcgi_data, fcgi_mark, fcgi_free_req, data);
The second memory leak is found in the
fcgi_stream_read()
function. It allocates a buffer to use when reading data from the stream,
but that buffer is not freed at any of the points where the function can be
exited. The fix is just to add
free(buff);
before each of the possible exit points.
With these two sets of changes in place, I have now ran more than two
million requests through FCGI, with a completely flat memory utilization. A
patch file with these changes can be found at
http://enigo.com/downloads/fcgi.c.patch
Thanks,
Kirk Haines
badly. I have a patch file at http://enigo.com/downloads/fcgi.c.patch if
you want to skip the explanation below and go right to it.
-----
http://enigo.com/projects/iowa/fcgipatch.html
When using the 0.8.5 version of the Ruby FCGI module with the C extension
instead of the pure Ruby extension, I noticed a substantial memory leak, to
the tune of around 16k+/request. A little poking through the code turned up
two seperate regions where memory is allocated without any provision to
free it later.
The first such instance occurs in the
fcgi_s_accept()
function. Memory is allocated to hold the FCGI request structure
(FCGX_Request), with this line:
req = ALLOC(FCGX_Request);
Then a data struct (fcgi_data), which has slots to hold the FCGI request,
the data streams, and the environment variables, is turned into a Ruby
object with a call to
Data_Make_Struct
A mark function is provided that marks the stream and environment slots for
the garbage collector. The request slot, however, is not marked since it
points to a C struct and not to a Ruby object. However, a free function is
not provided to cleanup either the root object or the request structure
that it contains. This is the first memory leak. It is fixed by providing a
function that will free this memory. This bug, with the same fix, was also
mentioned on ruby-core
(http://www.ruby-talk.org/cgi-bin/scat.rb/ruby/ruby-core/3245)last year in
reference to version 0.8.4.
Add this:
static void fcgi_free_req(fcgi_data *data)
{
free(data-<req);
free(data);
}
And then replace this:
obj = Data_Make_Struct(self, fcgi_data, fcgi_mark, 0, data);
with this:
obj = Data_Make_Struct(self, fcgi_data, fcgi_mark, fcgi_free_req, data);
The second memory leak is found in the
fcgi_stream_read()
function. It allocates a buffer to use when reading data from the stream,
but that buffer is not freed at any of the points where the function can be
exited. The fix is just to add
free(buff);
before each of the possible exit points.
With these two sets of changes in place, I have now ran more than two
million requests through FCGI, with a completely flat memory utilization. A
patch file with these changes can be found at
http://enigo.com/downloads/fcgi.c.patch
Thanks,
Kirk Haines