Re: crash in pipe_create()

[enami tsugutomo <enami%sm.sony.co.jp@localhost>]

Matthias Drochner <M.Drochner%fz-juelich.de@localhost> writes:

> enami%sm.sony.co.jp@localhost said:
> > pool_cachge_get() returned NULL, doesn't it?  It looks previous
> > revsion of pipe_create() returns with error on such case.
> 
> I should add that that box could sustain that workload
> with ease a week before. Also see my other backtrace - the
> system was paging and called pool_drain(). This shouldn't
> happen, it has 1G of RAM.
> There is something seriously wrong.

The allocation I mentioned above (the one in pipe_ctor) requests
pageable memory.  So, amount of physical memory doesn't matter.  (and
the code is committed this week).

Since the newly introduced pipe_ctor returns failure, pool_cache_get()
has to check if the returned value isn't NULL even if it passes
PR_WAIT.  Or, since we are allowed to wait, let uvm_km_alloc to wait
also.

Of course, there might be yet another something seriously wrong.

enami.

Index: sys_pipe.c
===================================================================
RCS file: /cvsroot/src/sys/kern/sys_pipe.c,v
retrieving revision 1.106
diff -u -r1.106 sys_pipe.c
--- sys_pipe.c  1 Feb 2009 18:23:04 -0000       1.106
+++ sys_pipe.c  6 Feb 2009 09:18:53 -0000
@@ -199,10 +199,9 @@
        memset(pipe, 0, sizeof(struct pipe));
        if (arg != NULL) {
                /* Preallocate space. */
-               va = uvm_km_alloc(kernel_map, PIPE_SIZE, 0, UVM_KMF_PAGEABLE);
-               if (va == 0) {
-                       return ENOMEM;
-               }
+               va = uvm_km_alloc(kernel_map, PIPE_SIZE, 0,
+                   UVM_KMF_PAGEABLE | UVM_KMF_WAITVA);
+               KASSERT(va != 0);
                pipe->pipe_kmem = va;
                atomic_add_int(&amountpipekva, PIPE_SIZE);
        }