Re: [Myricom help #6325] --disable-directcopy in mpich

[email protected] said:
> Yup - the local_mutex bug is fixed in the current source.
> gm-1.4pre51 is tagged as of yesterday and
> gm-1.4 will happen today (if I can get it out).

Not here.  Grabbed cvs after your message, made sure it got rebuilt with
--enable-directcopy this time, stuck in and configured the module, ran the
test case and immediately got one node with 4 processes all stock in
down().

Second issue today.  Back to --disable-directcopy.  With a different
test code which causes heavy swapping on the nodes, I'll eventually get
an oops (again randomly it seems) which can be traced back to the call
to free_page in gm_arch_unlock_user_buffer_page.  I put a check to see
if the page was in the swap cache (which is what makes linux
__free_pages_ok do BUG), and it always and only gets printed immediately
before that oops.  Are you sure you can call free_pages() on a highmem
page?

But later I managed to get another oops, this time from a bad pointer
0x5886b004 in:
    _gm_hash_reference_ptr
    _gm_hash_where
    gm_hash_insert
    gm_add_mapping_to_page_table

Perhaps I shouldn't be expecting gm to work with linux 2.4.0 with
highmem enabled?

		-- Pete

Index: drivers/linux/gm/gm_arch.c
===================================================================
RCS file: /ufs/src/repositories/gm/drivers/linux/gm/gm_arch.c,v
retrieving revision 1.205
diff -u -r1.205 gm_arch.c
--- drivers/linux/gm/gm_arch.c	2001/02/13 21:03:37	1.205
+++ drivers/linux/gm/gm_arch.c	2001/02/16 18:38:25
@@ -1309,6 +1309,9 @@
 		GM_WARN(("gm_arch_unlock_user_buffer_page: page=%p page_address is NULL\n",page));
 	}
 	else {
+	  if (PageSwapCache(page)) {
+	      GM_WARN(("gm_arch_unlock_user_buffer_page: page=%p on swap cache --pw\n", page));
+	  } else
 	  free_page(page_addr);
 #if LINUX_XX >= 24
 	  kunmap(page);