5 files changed, 35 insertions, 10 deletions

diff --git a/vlib/vlib/buffer.h b/vlib/vlib/buffer.h
index 9c148ef2f21..07ed85d8c31 100644
--- a/vlib/vlib/buffer.h
+++ b/vlib/vlib/buffer.h
@@ -59,7 +59,7 @@
 #ifdef CLIB_HAVE_VEC128
 typedef u8x16 vlib_copy_unit_t;
 #else
-typedef uword vlib_copy_unit_t;
+typedef u64 vlib_copy_unit_t;
 #endif
 
 /** \file
diff --git a/vlib/vlib/threads.c b/vlib/vlib/threads.c
index a69e4555af0..efbef379ba9 100644
--- a/vlib/vlib/threads.c
+++ b/vlib/vlib/threads.c
@@ -12,16 +12,18 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
+#define _GNU_SOURCE
+#include <sched.h>
+
 #include <signal.h>
 #include <math.h>
 #include <vppinfra/format.h>
 #include <vlib/vlib.h>
 
 #include <vlib/threads.h>
-#include <vlib/unix/physmem.h>
-
 #include <vlib/unix/cj.h>
 
+
 #if DPDK==1
 #include <rte_config.h>
 #include <rte_common.h>
@@ -175,6 +177,16 @@ vlib_thread_init (vlib_main_t * vm)
   if (!tm->cpu_socket_bitmap)
     tm->cpu_socket_bitmap = clib_bitmap_set(0, 0, 1);
 
+  /* pin main thread to main_lcore  */
+#if DPDK==0
+  {
+     cpu_set_t cpuset;
+     CPU_ZERO(&cpuset);
+     CPU_SET(tm->main_lcore, &cpuset);
+     pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
+  }
+#endif
+
   /* as many threads as stacks... */
   vec_validate_aligned (vlib_worker_threads, vec_len(vlib_thread_stacks)-1,
                         CLIB_CACHE_LINE_BYTES);
@@ -486,7 +498,6 @@ void *vlib_worker_thread_bootstrap_fn (void *arg)
 static int
 vlib_launch_thread (void *fp, vlib_worker_thread_t *w, unsigned lcore_id)
 {
-  pthread_t dummy;
   void *(*fp_arg)(void *) = fp;
 
 #if DPDK==1
@@ -497,7 +508,19 @@ vlib_launch_thread (void *fp, vlib_worker_thread_t *w, unsigned lcore_id)
       return -1;
   else
 #endif
-    return pthread_create (&dummy, NULL /* attr */, fp_arg, (void *)w);
+  {
+    int ret;
+    pthread_t worker;
+    cpu_set_t cpuset;
+    CPU_ZERO(&cpuset);
+    CPU_SET(lcore_id, &cpuset);
+
+    ret = pthread_create (&worker, NULL /* attr */, fp_arg, (void *)w);
+    if(ret == 0)
+        return pthread_setaffinity_np(worker, sizeof(cpu_set_t), &cpuset);
+    else
+        return ret;
+  }
 }
 
 static clib_error_t * start_workers (vlib_main_t * vm)
@@ -1031,7 +1054,7 @@ cpu_config (vlib_main_t * vm, unformat_input_t * input)
 
 VLIB_EARLY_CONFIG_FUNCTION (cpu_config, "cpu");
 
-#if !defined (__x86_64__) && !defined (__aarch64__) && !defined (__powerpc64__)
+#if !defined (__x86_64__) && !defined (__aarch64__) && !defined (__powerpc64__) && !defined(__arm__)
 void __sync_fetch_and_add_8 (void)
 {
   fformat(stderr, "%s called\n", __FUNCTION__);
diff --git a/vlib/vlib/threads.h b/vlib/vlib/threads.h
index 5773ed05a62..ce93b2c67cb 100644
--- a/vlib/vlib/threads.h
+++ b/vlib/vlib/threads.h
@@ -48,7 +48,9 @@ typedef struct vlib_thread_registration_ {
  * Make VLIB_MAX_CPUS a power-of-two, please...
  */
 
+#ifndef VLIB_MAX_CPUS
 #define VLIB_MAX_CPUS 256
+#endif
 
 #if VLIB_MAX_CPUS > CLIB_MAX_MHEAPS
 #error Please increase number of per-cpu mheaps
diff --git a/vlib/vlib/unix/cj.c b/vlib/vlib/unix/cj.c
index 665a13fa4f5..782ddce3099 100644
--- a/vlib/vlib/unix/cj.c
+++ b/vlib/vlib/unix/cj.c
@@ -40,8 +40,8 @@ cj_log (u32 type, void * data0, void * data1)
   r->time = vlib_time_now (cjm->vlib_main);
   r->cpu = os_get_cpu_number();
   r->type = type;
-  r->data[0] = (u64) data0;
-  r->data[1] = (u64) data1;
+  r->data[0] = pointer_to_uword(data0);
+  r->data[1] = pointer_to_uword(data1);
 }
 
 void cj_stop(void)
diff --git a/vlib/vlib/unix/physmem.c b/vlib/vlib/unix/physmem.c
index d7428c9b227..185483d6f14 100644
--- a/vlib/vlib/unix/physmem.c
+++ b/vlib/vlib/unix/physmem.c
@@ -158,10 +158,10 @@ static int htlb_init (vlib_main_t * vm)
      /* Don't want mheap mmap/munmap with IO memory. */
      MHEAP_FLAG_DISABLE_VM);
   
-  cur = (u64) pm->mem;
+  cur = pointer_to_uword(pm->mem);
   i = 0;
 
-  while (cur < (u64) pm->mem + pm->mem_size)
+  while (cur < pointer_to_uword(pm->mem) + pm->mem_size)
     {
       pfn = (u64) cur / pagesize;
       seek_loc = pfn * sizeof (u64);