svm: numa awareness for ssvm segments

Type: feature

Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I999836a7893a89aac5243b111eac35fddd03e2a6

4 files changed, 70 insertions, 26 deletions

diff --git a/src/svm/ssvm.c b/src/svm/ssvm.c
index 69463b1f192..66e38ebf864 100644
--- a/src/svm/ssvm.c
+++ b/src/svm/ssvm.c
@@ -90,6 +90,7 @@ ssvm_master_init_shm (ssvm_private_t * ssvm)
   mapa.requested_va = requested_va;
   mapa.size = ssvm->ssvm_size;
   mapa.fd = ssvm_fd;
+  mapa.numa_node = ssvm->numa;
   if (clib_mem_vm_ext_map (&mapa))
     {
       clib_unix_warning ("mmap");
@@ -237,6 +238,11 @@ ssvm_master_init_memfd (ssvm_private_t * memfd)
   alloc.size = memfd->ssvm_size;
   alloc.flags = CLIB_MEM_VM_F_SHARED;
   alloc.requested_va = memfd->requested_va;
+  if (memfd->numa)
+    {
+      alloc.numa_node = memfd->numa;
+      alloc.flags |= CLIB_MEM_VM_F_NUMA_PREFER;
+    }
   if ((err = clib_mem_vm_ext_alloc (&alloc)))
     {
       clib_error_report (err);
@@ -348,27 +354,31 @@ int
 ssvm_master_init_private (ssvm_private_t * ssvm)
 {
   uword pagesize = clib_mem_get_page_size (), rnd_size = 0;
+  clib_mem_vm_alloc_t alloc = { 0 };
+  struct dlmallinfo dlminfo;
   ssvm_shared_header_t *sh;
+  clib_error_t *err;
   u8 *heap;
 
   rnd_size = clib_max (ssvm->ssvm_size + (pagesize - 1), ssvm->ssvm_size);
   rnd_size &= ~(pagesize - 1);
 
-#if USE_DLMALLOC == 0
-  {
-    mheap_t *heap_header;
-
-    heap = mheap_alloc (0, rnd_size);
-    if (heap == 0)
-      {
-	clib_unix_warning ("mheap alloc");
-	return -1;
-      }
-    heap_header = mheap_header (heap);
-    heap_header->flags |= MHEAP_FLAG_THREAD_SAFE;
-  }
-#else
-  heap = create_mspace (rnd_size, 1 /* locked */ );
+  alloc.name = (char *) ssvm->name;
+  alloc.size = rnd_size + pagesize;
+  if (ssvm->numa)
+    {
+      alloc.numa_node = ssvm->numa;
+      alloc.flags |= CLIB_MEM_VM_F_NUMA_PREFER;
+    }
+
+  if ((err = clib_mem_vm_ext_alloc (&alloc)))
+    {
+      clib_error_report (err);
+      return SSVM_API_ERROR_CREATE_FAILURE;
+    }
+
+  heap = create_mspace_with_base ((u8 *) alloc.addr + pagesize, rnd_size,
+				  1 /* locked */ );
   if (heap == 0)
     {
       clib_unix_warning ("mheap alloc");
@@ -378,18 +388,16 @@ ssvm_master_init_private (ssvm_private_t * ssvm)
   mspace_disable_expand (heap);
 
   /* Find actual size because mspace size is rounded up by dlmalloc */
-  struct dlmallinfo dlminfo;
   dlminfo = mspace_mallinfo (heap);
   rnd_size = dlminfo.fordblks;
-#endif
 
   ssvm->ssvm_size = rnd_size;
   ssvm->i_am_master = 1;
   ssvm->my_pid = getpid ();
   ssvm->requested_va = ~0;
 
-  /* Allocate a [sic] shared memory header, in process memory... */
-  sh = clib_mem_alloc_aligned (sizeof (*sh), CLIB_CACHE_LINE_BYTES);
+  /* First page in allocated memory is set aside for the shared header */
+  sh = alloc.addr;
   ssvm->sh = sh;
 
   clib_memset (sh, 0, sizeof (*sh));
@@ -413,12 +421,8 @@ void
 ssvm_delete_private (ssvm_private_t * ssvm)
 {
   vec_free (ssvm->name);
-#if USE_DLMALLOC == 0
-  mheap_free (ssvm->sh->heap);
-#else
   destroy_mspace (ssvm->sh->heap);
-#endif
-  clib_mem_free (ssvm->sh);
+  clib_mem_vm_free (ssvm->sh, ssvm->ssvm_size + clib_mem_get_page_size ());
 }
 
 int
diff --git a/src/svm/ssvm.h b/src/svm/ssvm.h
index 82951d614b0..6225a63cae4 100644
--- a/src/svm/ssvm.h
+++ b/src/svm/ssvm.h
@@ -83,9 +83,10 @@ typedef struct
 {
   ssvm_shared_header_t *sh;
   uword ssvm_size;
+  uword requested_va;
   u32 my_pid;
   u8 *name;
-  uword requested_va;
+  u8 numa;			/**< Numa requested at alloc time */
   int i_am_master;
 
   union
diff --git a/src/vppinfra/linux/mem.c b/src/vppinfra/linux/mem.c
index 69ac871567c..3b7294f038a 100644
--- a/src/vppinfra/linux/mem.c
+++ b/src/vppinfra/linux/mem.c
@@ -386,21 +386,59 @@ done:
 clib_error_t *
 clib_mem_vm_ext_map (clib_mem_vm_map_t * a)
 {
+  long unsigned int old_mask[16] = { 0 };
   int mmap_flags = MAP_SHARED;
+  clib_error_t *err = 0;
+  int old_mpol = -1;
   void *addr;
+  int rv;
+
+  if (a->numa_node)
+    {
+      rv = get_mempolicy (&old_mpol, old_mask, sizeof (old_mask) * 8 + 1, 0,
+			  0);
+
+      if (rv == -1)
+	{
+	  err = clib_error_return_unix (0, "get_mempolicy");
+	  goto done;
+	}
+    }
 
   if (a->requested_va)
     mmap_flags |= MAP_FIXED;
 
+  if (old_mpol != -1)
+    {
+      long unsigned int mask[16] = { 0 };
+      mask[0] = 1 << a->numa_node;
+      rv = set_mempolicy (MPOL_BIND, mask, sizeof (mask) * 8 + 1);
+      if (rv == -1)
+	{
+	  err = clib_error_return_unix (0, "set_mempolicy");
+	  goto done;
+	}
+    }
+
   addr = (void *) mmap (uword_to_pointer (a->requested_va, void *), a->size,
 			PROT_READ | PROT_WRITE, mmap_flags, a->fd, 0);
 
   if (addr == MAP_FAILED)
     return clib_error_return_unix (0, "mmap");
 
+  /* re-apply old numa memory policy */
+  if (old_mpol != -1 &&
+      set_mempolicy (old_mpol, old_mask, sizeof (old_mask) * 8 + 1) == -1)
+    {
+      err = clib_error_return_unix (0, "set_mempolicy");
+      goto done;
+    }
+
   a->addr = addr;
   CLIB_MEM_UNPOISON (addr, a->size);
-  return 0;
+
+done:
+  return err;
 }
 
 /*
diff --git a/src/vppinfra/mem.h b/src/vppinfra/mem.h
index 0367c4a1213..f35c495eb25 100644
--- a/src/vppinfra/mem.h
+++ b/src/vppinfra/mem.h
@@ -455,6 +455,7 @@ typedef struct
   int fd;		/**< File descriptor to be mapped */
   uword requested_va;	/**< Request fixed position mapping */
   void *addr;		/**< Pointer to mapped memory, if successful */
+  u8 numa_node;
 } clib_mem_vm_map_t;
 
 clib_error_t *clib_mem_vm_ext_map (clib_mem_vm_map_t * a);