|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [Xen-devel] Re: [PATCH 1 of 2] Fix invalid memory access in OCaml mmap library (to play nicely with the GC)
On 30/09/2011 17:34, Ian Campbell wrote: On Fri, 2011-09-30 at 16:39 +0100, Zheng Li wrote:This was a bug reported by Roberto Di Cosmo. When he tried to reuse the mmap library for his own project, Mmap.read occasionally got different result when reading from the same map. This turned out to be a bug in the binding, where a C pointer was created pointing to a OCaml value, and the OCaml value was subsequently moved around by the GC after memory allocation and hence invalidated the C pointer. This patch removes the indirection of C pointer and uses OCaml macro to access values directly.I was initially quite confused about how the value of a parameter to a C function could change while that function was running but I see now that the CAMLparam* stuff apparently enables that behaviour by stashing the address of the parameters so if you ever call back into the runtime (and presumably asynchronously if you are multithreaded) the GC can indeed move stuff around.
I think the issue itself is not related to multi-thread though. In the original
version of the stub_mmap_read,
intf = GET_C_STRUCT(interface);
"interface" was passed to the C function as an OCaml value registered in the OCaml runtime, underline it was just a
pointer pointing to a block in the OCaml heap. It was copied and casted to the "intf" on the C side, hence
"intf" was initially pointing to the same address as "interface". Unfortunately, before we made any use of
"intf", we first called
data = caml_alloc_string(c_len);
which called back to the OCaml runtime and allocated memory in the OCaml heap. Since caml_alloc_string was a
function defined by OCaml runtime, it deliberately called quite a few GC routines under the hood which can
possibly move any values in the OCaml heap. OCaml runtime would certainly update the "interface"
pointer accordingly since it was registered with the runtime. But "intf" was a hard copy on the C
side that OCaml had no idea about. So "intf" still pointed to the old address which was no longer
valid. Despite the back and forth between C and OCaml territories, all was single threaded here.
The proposed patch only appears to fix the issue if there is some higher level locking which prevents another thread from triggering a gc while this function is running. IIRC there is a single global lock which covers all C code called from ocaml, is that right?
OCaml runtime has a global lock that will only allow one thread to execute it
at a time and the GC is single threaded as well. This ensures the safety but
unfortunately introduces the bottleneck. On the other hand, for threads that
don't need to access memory on the OCaml side for a while (e.g. long running
pure C routines invoked from OCaml side), they can run in parallel by using
caml_{enter|leaving}_blocking_section primitives to give up and then retain the
rights to enter OCaml runtime.
Cheers
--
Zheng
_______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxxxxxxxx http://lists.xensource.com/xen-devel
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |