5 files changed, 35 insertions, 11 deletions

diff --git a/vlib/vlib/main.h b/vlib/vlib/main.h
index 5a8d745661b..1a110459826 100644
--- a/vlib/vlib/main.h
+++ b/vlib/vlib/main.h
@@ -171,6 +171,7 @@ typedef struct vlib_main_t {
   /* control-plane API queue signal pending */
   volatile u32 queue_signal_pending;
   void (*queue_signal_callback)(struct vlib_main_t *);
+  u8 **argv;
 } vlib_main_t;
 
 /* Global main structure. */
diff --git a/vlib/vlib/node.c b/vlib/vlib/node.c
index 4fb117e4f3e..66a2d09fdb8 100644
--- a/vlib/vlib/node.c
+++ b/vlib/vlib/node.c
@@ -262,7 +262,7 @@ static void node_elog_init (vlib_main_t * vm, uword ni)
 }
 
 #ifdef CLIB_UNIX
-#define STACK_ALIGN 4096
+#define STACK_ALIGN (clib_mem_get_page_size())
 #else
 #define STACK_ALIGN CLIB_CACHE_LINE_BYTES
 #endif
@@ -272,6 +272,7 @@ static void register_node (vlib_main_t * vm,
 {
   vlib_node_main_t * nm = &vm->node_main;
   vlib_node_t * n;
+  u32 page_size = clib_mem_get_page_size();
   int i;
 
   if (CLIB_DEBUG > 0)
@@ -363,9 +364,26 @@ static void register_node (vlib_main_t * vm,
 
 	log2_n_stack_bytes = clib_max (r->process_log2_n_stack_bytes, 15);
 
-	p = clib_mem_alloc_aligned_no_fail 
+#ifdef CLIB_UNIX
+        /* 
+         * Bump the stack size if running over a kernel with a large page size,
+         * and the stack isn't any too big to begin with. Otherwise, we'll
+         * trip over the stack guard page for sure.
+         */
+        if ((page_size > (4<<10)) && log2_n_stack_bytes < 19)
+          {
+            if ((1<<log2_n_stack_bytes) <= page_size)
+              log2_n_stack_bytes = min_log2 (page_size) + 1;
+            else
+              log2_n_stack_bytes++;
+          }
+#endif
+
+	p = clib_mem_alloc_aligned_at_offset 
             (sizeof (p[0]) + (1 << log2_n_stack_bytes),
-             STACK_ALIGN);
+             STACK_ALIGN, STRUCT_OFFSET_OF (vlib_process_t, stack));
+        if (p == 0)
+            clib_panic ("failed to allocate process stack (%d bytes)", 1<<log2_n_stack_bytes);
 
 	memset (p, 0, sizeof (p[0]));
 	p->log2_n_stack_bytes = log2_n_stack_bytes;
@@ -388,7 +406,7 @@ static void register_node (vlib_main_t * vm,
          * Disallow writes to the bottom page of the stack, to
          * catch stack overflows.
          */
-        if (mprotect (p->stack, 4096, PROT_READ) < 0)
+        if (mprotect (p->stack, page_size, PROT_READ) < 0)
             clib_unix_warning ("process stack");
 #endif
 
diff --git a/vlib/vlib/threads.c b/vlib/vlib/threads.c
index 73abba83f70..3eed1080d31 100644
--- a/vlib/vlib/threads.c
+++ b/vlib/vlib/threads.c
@@ -1025,7 +1025,7 @@ cpu_config (vlib_main_t * vm, unformat_input_t * input)
 
 VLIB_EARLY_CONFIG_FUNCTION (cpu_config, "cpu");
 
-#if !defined (__x86_64__)
+#if !defined (__x86_64__) && !defined (__aarch64__) && !defined (__powerpc64__)
 void __sync_fetch_and_add_8 (void)
 {
   fformat(stderr, "%s called\n", __FUNCTION__);
@@ -1128,6 +1128,7 @@ show_threads_fn (vlib_main_t * vm,
                    "ID", "Name", "Type", "LWP",
                    "lcore", "Core", "Socket", "State");
 
+#if !defined(__powerpc64__)
   for (i = 0; i < vec_len(vlib_worker_threads); i++)
     {
       w = vlib_worker_threads + i;
@@ -1166,6 +1167,7 @@ show_threads_fn (vlib_main_t * vm,
       vlib_cli_output(vm, "%v", line);
       vec_free(line);
     }
+#endif
 
   return 0;
 }
diff --git a/vlib/vlib/unix/main.c b/vlib/vlib/unix/main.c
index b85f3e73326..1aac32759a8 100644
--- a/vlib/vlib/unix/main.c
+++ b/vlib/vlib/unix/main.c
@@ -397,15 +397,13 @@ VLIB_MAIN_LOOP_EXIT_FUNCTION (unix_exit);
 
 u8 **vlib_thread_stacks;
 
-static char **argv_global;
-  
 static uword thread0 (uword arg)
 {
   vlib_main_t * vm = (vlib_main_t *)arg;
   unformat_input_t input;
   int i;
   
-  unformat_init_command_line (&input, argv_global);
+  unformat_init_command_line (&input, vm->argv);
   i = vlib_main (vm, &input);
   unformat_free (&input);
   
@@ -423,7 +421,7 @@ int vlib_unix_main (int argc, char * argv[])
   clib_error_t * e;
   int i;
 
-  argv_global = argv;
+  vm->argv = (u8 **)argv;
   vm->name = argv[0];
   vm->heap_base = clib_mem_get_heap ();
   ASSERT(vm->heap_base);
@@ -432,7 +430,7 @@ int vlib_unix_main (int argc, char * argv[])
   if (i)
     return i;
   
-  unformat_init_command_line (&input, argv_global);
+  unformat_init_command_line (&input, vm->argv);
   vm->init_functions_called = hash_create (0, /* value bytes */ 0);
   e = vlib_call_all_config_functions (vm, &input, 1 /* early */);
   if (e != 0)
@@ -459,7 +457,7 @@ int vlib_unix_main (int argc, char * argv[])
        * Disallow writes to the bottom page of the stack, to
        * catch stack overflows.
        */
-      if (mprotect (thread_stacks, 4096, PROT_READ) < 0)
+      if (mprotect (thread_stacks, clib_mem_get_page_size(), PROT_READ) < 0)
           clib_unix_warning ("thread stack");
 
       thread_stacks += VLIB_THREAD_STACK_SIZE;
diff --git a/vlib/vlib/unix/physmem.c b/vlib/vlib/unix/physmem.c
index 83b40be6449..d7428c9b227 100644
--- a/vlib/vlib/unix/physmem.c
+++ b/vlib/vlib/unix/physmem.c
@@ -261,6 +261,11 @@ clib_error_t * unix_physmem_init (vlib_main_t * vm, int physical_memory_required
   else
     error = clib_error_return (0, "uio_dma deprecated");
 
+  vpm->page_mask = pow2_mask (vpm->log2_n_bytes_per_page);
+  vpm->virtual.start = pointer_to_uword (pm->mem);
+  vpm->virtual.size = pm->mem_size;
+  vpm->virtual.end = vpm->virtual.start + vpm->virtual.size;
+
   if (using_fake_memory)
       fformat(stderr, "%s: use fake dma pages\n", __FUNCTION__);
   else