stats: configurable page size

Type: improvement
Change-Id: I9973bce20a0a2a8a7e227cf96518de5b79374425
Signed-off-by: Damjan Marion <damarion@cisco.com>

7 files changed, 163 insertions, 91 deletions

diff --git a/src/vpp/conf/startup.conf b/src/vpp/conf/startup.conf
index 12679da778c..0be10d5ba19 100644
--- a/src/vpp/conf/startup.conf
+++ b/src/vpp/conf/startup.conf
@@ -186,6 +186,7 @@ cpu {
     # socket-name <filename>, name of the stats segment socket
     #     defaults to /run/vpp/stats.sock
     # size <nnn>[KMG], size of the stats segment, defaults to 32mb
+    # page-size <nnn>, page size, ie. 2m, defaults to 4k
     # per-node-counters on | off, defaults to none
     # update-interval <f64-seconds>, sets the segment scrape / update interval
 # }
diff --git a/src/vpp/stats/stat_segment.c b/src/vpp/stats/stat_segment.c
index 8255b16c010..968c0566b3f 100644
--- a/src/vpp/stats/stat_segment.c
+++ b/src/vpp/stats/stat_segment.c
@@ -316,32 +316,37 @@ vlib_map_stat_segment_init (void)
   stat_segment_main_t *sm = &stat_segment_main;
   stat_segment_shared_header_t *shared_header;
   void *oldheap;
-  ssize_t memory_size;
+  uword memory_size, sys_page_sz;
   int mfd;
-  char *mem_name = "stat_segment_test";
-  void *memaddr;
+  char *mem_name = "stat segment";
+  void *heap, *memaddr;
 
   memory_size = sm->memory_size;
   if (memory_size == 0)
     memory_size = STAT_SEGMENT_DEFAULT_SIZE;
 
-  /* Create shared memory segment */
-  if ((mfd = memfd_create (mem_name, 0)) < 0)
-    return clib_error_return (0, "stat segment memfd_create failure");
+  if (sm->log2_page_sz == CLIB_MEM_PAGE_SZ_UNKNOWN)
+    sm->log2_page_sz = CLIB_MEM_PAGE_SZ_DEFAULT;
+
+  mfd = clib_mem_vm_create_fd (sm->log2_page_sz, mem_name);
 
   /* Set size */
   if ((ftruncate (mfd, memory_size)) == -1)
     return clib_error_return (0, "stat segment ftruncate failure");
 
-  if ((memaddr =
-       mmap (NULL, memory_size, PROT_READ | PROT_WRITE, MAP_SHARED, mfd,
-	     0)) == MAP_FAILED)
+  if (mfd == -1)
+    return clib_error_return (0, "stat segment memory fd failure: %U",
+			      format_clib_error, clib_mem_get_last_error ());
+
+  memaddr = clib_mem_vm_map_shared (0, memory_size, mfd, 0, mem_name);
+
+  if (memaddr == CLIB_MEM_VM_MAP_FAILED)
     return clib_error_return (0, "stat segment mmap failure");
 
-  void *heap;
-  heap =
-    create_mspace_with_base (((u8 *) memaddr) + getpagesize (),
-			     memory_size - getpagesize (), 1 /* locked */ );
+  sys_page_sz = clib_mem_get_page_size ();
+
+  heap = create_mspace_with_base (((u8 *) memaddr) + sys_page_sz, memory_size
+				  - sys_page_sz, 1 /* locked */ );
   mspace_disable_expand (heap);
   sm->heap = heap;
   sm->memfd = mfd;
@@ -904,6 +909,9 @@ statseg_config (vlib_main_t * vm, unformat_input_t * input)
       else if (unformat (input, "size %U",
 			 unformat_memory_size, &sm->memory_size))
 	;
+      else if (unformat (input, "page-size %U",
+			 unformat_log2_page_size, &sm->log2_page_sz))
+	;
       else if (unformat (input, "per-node-counters on"))
 	sm->node_counters_enabled = 1;
       else if (unformat (input, "per-node-counters off"))
diff --git a/src/vpp/stats/stat_segment.h b/src/vpp/stats/stat_segment.h
index a88067154f7..a048fa5b8ca 100644
--- a/src/vpp/stats/stat_segment.h
+++ b/src/vpp/stats/stat_segment.h
@@ -93,6 +93,7 @@ typedef struct
   clib_socket_t *socket;
   u8 *socket_name;
   ssize_t memory_size;
+  clib_mem_page_sz_t log2_page_sz;
   u8 node_counters_enabled;
   void *last;
   void *heap;
diff --git a/src/vppinfra/linux/mem.c b/src/vppinfra/linux/mem.c
index ef98b3f360b..76195a21118 100644
--- a/src/vppinfra/linux/mem.c
+++ b/src/vppinfra/linux/mem.c
@@ -54,6 +54,10 @@
 #define MAP_HUGE_SHIFT 26
 #endif
 
+#ifndef MFD_HUGE_SHIFT
+#define MFD_HUGE_SHIFT 26
+#endif
+
 #ifndef MAP_FIXED_NOREPLACE
 #define MAP_FIXED_NOREPLACE 0x100000
 #endif
@@ -193,75 +197,9 @@ clib_mem_vm_randomize_va (uword * requested_va,
 }
 
 clib_error_t *
-clib_mem_create_fd (char *name, int *fdp)
-{
-  int fd;
-
-  ASSERT (name);
-
-  if ((fd = memfd_create (name, MFD_ALLOW_SEALING)) == -1)
-    return clib_error_return_unix (0, "memfd_create");
-
-  if ((fcntl (fd, F_ADD_SEALS, F_SEAL_SHRINK)) == -1)
-    {
-      close (fd);
-      return clib_error_return_unix (0, "fcntl (F_ADD_SEALS)");
-    }
-
-  *fdp = fd;
-  return 0;
-}
-
-clib_error_t *
-clib_mem_create_hugetlb_fd (char *name, int *fdp)
-{
-  clib_error_t *err = 0;
-  int fd = -1;
-  static int memfd_hugetlb_supported = 1;
-  char *mount_dir;
-  char template[] = "/tmp/hugepage_mount.XXXXXX";
-  u8 *filename;
-
-  ASSERT (name);
-
-  if (memfd_hugetlb_supported)
-    {
-      if ((fd = memfd_create (name, MFD_HUGETLB)) != -1)
-	goto done;
-
-      /* avoid further tries if memfd MFD_HUGETLB is not supported */
-      if (errno == EINVAL && strnlen (name, 256) <= 249)
-	memfd_hugetlb_supported = 0;
-    }
-
-  mount_dir = mkdtemp (template);
-  if (mount_dir == 0)
-    return clib_error_return_unix (0, "mkdtemp \'%s\'", template);
-
-  if (mount ("none", (char *) mount_dir, "hugetlbfs", 0, NULL))
-    {
-      rmdir ((char *) mount_dir);
-      err = clib_error_return_unix (0, "mount hugetlb directory '%s'",
-				    mount_dir);
-    }
-
-  filename = format (0, "%s/%s%c", mount_dir, name, 0);
-  fd = open ((char *) filename, O_CREAT | O_RDWR, 0755);
-  umount2 ((char *) mount_dir, MNT_DETACH);
-  rmdir ((char *) mount_dir);
-
-  if (fd == -1)
-    err = clib_error_return_unix (0, "open");
-
-done:
-  if (fd != -1)
-    fdp[0] = fd;
-  return err;
-}
-
-clib_error_t *
 clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a)
 {
+  clib_mem_main_t *mm = &clib_mem_main;
   int fd = -1;
   clib_error_t *err = 0;
   void *addr = 0;
@@ -301,15 +239,16 @@ clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a)
       /* if hugepages are needed we need to create mount point */
       if (a->flags & CLIB_MEM_VM_F_HUGETLB)
 	{
-	  if ((err = clib_mem_create_hugetlb_fd (a->name, &fd)))
-	    goto error;
-
+	  log2_page_size = CLIB_MEM_PAGE_SZ_DEFAULT_HUGE;
 	  mmap_flags |= MAP_LOCKED;
 	}
       else
+	log2_page_size = CLIB_MEM_PAGE_SZ_DEFAULT;
+
+      if ((fd = clib_mem_vm_create_fd (log2_page_size, "%s", a->name)) == -1)
 	{
-	  if ((err = clib_mem_create_fd (a->name, &fd)))
-	    goto error;
+	  err = clib_error_return (0, "%U", format_clib_error, mm->error);
+	  goto error;
 	}
 
       log2_page_size = clib_mem_get_fd_log2_page_size (fd);
@@ -414,6 +353,111 @@ clib_mem_vm_ext_free (clib_mem_vm_alloc_t * a)
     }
 }
 
+static int
+legacy_memfd_create (u8 * name)
+{
+  clib_mem_main_t *mm = &clib_mem_main;
+  int fd = -1;
+  char *mount_dir;
+  u8 *filename;
+
+  /* create mount directory */
+  if ((mount_dir = mkdtemp ("/tmp/hugepage_mount.XXXXXX")) == 0)
+    {
+      vec_reset_length (mm->error);
+      mm->error = clib_error_return_unix (mm->error, "mkdtemp");
+      return -1;
+    }
+
+  if (mount ("none", mount_dir, "hugetlbfs", 0, NULL))
+    {
+      rmdir ((char *) mount_dir);
+      vec_reset_length (mm->error);
+      mm->error = clib_error_return_unix (mm->error, "mount");
+      return -1;
+    }
+
+  filename = format (0, "%s/%s%c", mount_dir, name, 0);
+
+  if ((fd = open ((char *) filename, O_CREAT | O_RDWR, 0755)) == -1)
+    {
+      vec_reset_length (mm->error);
+      mm->error = clib_error_return_unix (mm->error, "mkdtemp");
+    }
+
+  umount2 ((char *) mount_dir, MNT_DETACH);
+  rmdir ((char *) mount_dir);
+  vec_free (filename);
+
+  return fd;
+}
+
+int
+clib_mem_vm_create_fd (clib_mem_page_sz_t log2_page_size, char *fmt, ...)
+{
+  clib_mem_main_t *mm = &clib_mem_main;
+  int fd;
+  unsigned int memfd_flags;
+  va_list va;
+  u8 *s = 0;
+
+  if (log2_page_size == mm->log2_page_sz)
+    log2_page_size = CLIB_MEM_PAGE_SZ_DEFAULT;
+
+  switch (log2_page_size)
+    {
+    case CLIB_MEM_PAGE_SZ_UNKNOWN:
+      return -1;
+    case CLIB_MEM_PAGE_SZ_DEFAULT:
+      memfd_flags = MFD_ALLOW_SEALING;
+      break;
+    case CLIB_MEM_PAGE_SZ_DEFAULT_HUGE:
+      memfd_flags = MFD_HUGETLB;
+      break;
+    default:
+      memfd_flags = MFD_HUGETLB | log2_page_size << MFD_HUGE_SHIFT;
+    }
+
+  va_start (va, fmt);
+  s = va_format (0, fmt, &va);
+  va_end (va);
+
+  /* memfd_create maximum string size is 249 chars without trailing zero */
+  if (vec_len (s) > 249)
+    _vec_len (s) = 249;
+  vec_add1 (s, 0);
+
+  /* memfd_create introduced in kernel 3.17, we don't support older kernels */
+  fd = memfd_create ((char *) s, memfd_flags);
+
+  /* kernel versions < 4.14 does not support memfd_create for huge pages */
+  if (fd == -1 && errno == EINVAL &&
+      log2_page_size == CLIB_MEM_PAGE_SZ_DEFAULT_HUGE)
+    {
+      fd = legacy_memfd_create (s);
+    }
+  else if (fd == -1)
+    {
+      vec_reset_length (mm->error);
+      mm->error = clib_error_return_unix (mm->error, "memfd_create");
+      vec_free (s);
+      return -1;
+    }
+
+  vec_free (s);
+
+  if ((memfd_flags & MFD_ALLOW_SEALING) &&
+      ((fcntl (fd, F_ADD_SEALS, F_SEAL_SHRINK)) == -1))
+    {
+      vec_reset_length (mm->error);
+      mm->error = clib_error_return_unix (mm->error, "fcntl (F_ADD_SEALS)");
+      close (fd);
+      return -1;
+    }
+
+  return fd;
+}
+
 uword
 clib_mem_vm_reserve (uword start, uword size, clib_mem_page_sz_t log2_page_sz)
 {
diff --git a/src/vppinfra/mem.h b/src/vppinfra/mem.h
index 9d0aa0997fc..3ba20ad6d3e 100644
--- a/src/vppinfra/mem.h
+++ b/src/vppinfra/mem.h
@@ -113,6 +113,9 @@ typedef struct
 
   /* memory maps */
   clib_mem_vm_map_hdr_t *first_map, *last_map;
+
+  /* last error */
+  clib_error_t *error;
 } clib_mem_main_t;
 
 extern clib_mem_main_t clib_mem_main;
@@ -461,8 +464,7 @@ clib_mem_get_log2_default_hugepage_size ()
   return clib_mem_main.log2_default_hugepage_sz;
 }
 
-clib_error_t *clib_mem_create_fd (char *name, int *fdp);
-clib_error_t *clib_mem_create_hugetlb_fd (char *name, int *fdp);
+int clib_mem_vm_create_fd (clib_mem_page_sz_t log2_page_size, char *fmt, ...);
 clib_error_t *clib_mem_vm_ext_alloc (clib_mem_vm_alloc_t * a);
 void clib_mem_vm_ext_free (clib_mem_vm_alloc_t * a);
 uword clib_mem_get_fd_page_size (int fd);
@@ -545,6 +547,12 @@ clib_mem_page_bytes (clib_mem_page_sz_t log2_page_size)
   return 1 << clib_mem_log2_page_size_validate (log2_page_size);
 }
 
+static_always_inline clib_error_t *
+clib_mem_get_last_error (void)
+{
+  return clib_mem_main.error;
+}
+
 
 #include <vppinfra/error.h>	/* clib_panic */
 
diff --git a/src/vppinfra/mem_dlmalloc.c b/src/vppinfra/mem_dlmalloc.c
index 50dc57a60bd..10d3c61c77e 100644
--- a/src/vppinfra/mem_dlmalloc.c
+++ b/src/vppinfra/mem_dlmalloc.c
@@ -543,6 +543,19 @@ mheap_alloc_with_lock (void *memory, uword size, int locked)
     }
 }
 
+void *
+clib_mem_create_heap (void *base, uword size, char *fmt, ...)
+{
+  base = clib_mem_vm_map_internal (base, CLIB_MEM_PAGE_SZ_DEFAULT, size, -1,
+				   0, "str");
+
+  if (base == 0)
+    return 0;
+
+  create_mspace_with_base (base, size, 1 /* locked */ );
+  return base;
+}
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vppinfra/pmalloc.c b/src/vppinfra/pmalloc.c
index f6171dbf458..cb6c7e3ba3b 100644
--- a/src/vppinfra/pmalloc.c
+++ b/src/vppinfra/pmalloc.c
@@ -288,10 +288,7 @@ pmalloc_map_pages (clib_pmalloc_main_t * pm, clib_pmalloc_arena_t * a,
   if (a->flags & CLIB_PMALLOC_ARENA_F_SHARED_MEM)
     {
       mmap_flags |= MAP_SHARED;
-      if (a->log2_subpage_sz != clib_mem_get_log2_page_size ())
-	pm->error = clib_mem_create_hugetlb_fd ((char *) a->name, &a->fd);
-      else
-	pm->error = clib_mem_create_fd ((char *) a->name, &a->fd);
+      a->fd = clib_mem_vm_create_fd (a->log2_subpage_sz, "%s", a->name);
       if (a->fd == -1)
 	goto error;
       if ((ftruncate (a->fd, size)) == -1)