svm session: use new clib mem apis for segments

"sh memory map" now reports shared memory segments mapped by session
layer for applications.

Type: improvement

Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I7debdcd06c7728abfb5cf817a1ccc37de495472c

7 files changed, 91 insertions, 61 deletions

diff --git a/src/svm/fifo_segment.c b/src/svm/fifo_segment.c
index 5fb2758606e..a247929a058 100644
--- a/src/svm/fifo_segment.c
+++ b/src/svm/fifo_segment.c
@@ -1362,7 +1362,7 @@ format_fifo_segment_type (u8 * s, va_list * args)
   ssvm_segment_type_t st = ssvm_type (&sp->ssvm);
 
   if (st == SSVM_SEGMENT_PRIVATE)
-    s = format (s, "%s", "private-heap");
+    s = format (s, "%s", "private");
   else if (st == SSVM_SEGMENT_MEMFD)
     s = format (s, "%s", "memfd");
   else if (st == SSVM_SEGMENT_SHM)
@@ -1400,7 +1400,7 @@ format_fifo_segment (u8 * s, va_list * args)
 
   if (fs == 0)
     {
-      s = format (s, "%-15s%15s%15s%15s%15s%15s", "Name", "Type",
+      s = format (s, "%-20s%10s%15s%15s%15s%15s", "Name", "Type",
 		  "HeapSize (M)", "ActiveFifos", "FreeFifos", "Address");
       return s;
     }
@@ -1409,7 +1409,7 @@ format_fifo_segment (u8 * s, va_list * args)
   active_fifos = fifo_segment_num_fifos (fs);
   free_fifos = fifo_segment_num_free_fifos (fs);
 
-  s = format (s, "%-15v%15U%15llu%15u%15u%15llx", ssvm_name (&fs->ssvm),
+  s = format (s, "%-20v%10U%15llu%15u%15u%15llx", ssvm_name (&fs->ssvm),
 	      format_fifo_segment_type, fs, size >> 20ULL, active_fifos,
 	      free_fifos, address);
 
diff --git a/src/svm/ssvm.c b/src/svm/ssvm.c
index d8831ec1c48..191dbe571c7 100644
--- a/src/svm/ssvm.c
+++ b/src/svm/ssvm.c
@@ -29,7 +29,6 @@ int
 ssvm_master_init_shm (ssvm_private_t * ssvm)
 {
   int ssvm_fd;
-  clib_mem_vm_map_t mapa = { 0 };
   u8 junk = 0, *ssvm_filename;
   ssvm_shared_header_t *sh;
   uword page_size, requested_va = 0;
@@ -84,19 +83,18 @@ ssvm_master_init_shm (ssvm_private_t * ssvm)
       clib_mem_vm_randomize_va (&requested_va, min_log2 (page_size));
     }
 
-  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))
+  sh = clib_mem_vm_map_shared (uword_to_pointer (requested_va, void *),
+			       ssvm->ssvm_size, ssvm_fd, 0,
+			       (char *) ssvm->name);
+  if (sh == CLIB_MEM_VM_MAP_FAILED)
     {
       clib_unix_warning ("mmap");
       close (ssvm_fd);
       return SSVM_API_ERROR_MMAP;
     }
+
   close (ssvm_fd);
 
-  sh = mapa.addr;
   CLIB_MEM_UNPOISON (sh, sizeof (*sh));
   sh->master_pid = ssvm->my_pid;
   sh->ssvm_size = ssvm->ssvm_size;
@@ -206,7 +204,10 @@ ssvm_delete_shm (ssvm_private_t * ssvm)
   vec_free (fn);
   vec_free (ssvm->name);
 
-  munmap ((void *) ssvm->sh, ssvm->ssvm_size);
+  if (ssvm->i_am_master)
+    clib_mem_vm_unmap (ssvm->sh);
+  else
+    munmap ((void *) ssvm->sh, ssvm->ssvm_size);
 }
 
 /**
@@ -215,44 +216,60 @@ ssvm_delete_shm (ssvm_private_t * ssvm)
 int
 ssvm_master_init_memfd (ssvm_private_t * memfd)
 {
+  int log2_page_size, n_pages;
   uword page_size;
   ssvm_shared_header_t *sh;
   void *oldheap;
-  clib_mem_vm_alloc_t alloc = { 0 };
-  clib_error_t *err;
 
   if (memfd->ssvm_size == 0)
     return SSVM_API_ERROR_NO_SIZE;
 
   ASSERT (vec_c_string_is_terminated (memfd->name));
 
-  alloc.name = (char *) memfd->name;
-  alloc.size = memfd->ssvm_size;
-  alloc.flags = CLIB_MEM_VM_F_SHARED;
-  alloc.requested_va = memfd->requested_va;
-  if (memfd->numa)
+  memfd->fd = clib_mem_vm_create_fd (CLIB_MEM_PAGE_SZ_DEFAULT,
+				     (char *) memfd->name);
+
+  if (memfd->fd == CLIB_MEM_ERROR)
+    {
+      clib_unix_warning (0, "failed to create memfd");
+      return SSVM_API_ERROR_CREATE_FAILURE;
+    }
+
+  log2_page_size = clib_mem_get_fd_log2_page_size (memfd->fd);
+  if (log2_page_size == 0)
+    {
+      clib_unix_warning (0, "cannot determine page size");
+      return SSVM_API_ERROR_CREATE_FAILURE;
+    }
+
+  n_pages = ((memfd->ssvm_size - 1) >> log2_page_size) + 1;
+
+  if ((ftruncate (memfd->fd, n_pages << log2_page_size)) == -1)
     {
-      alloc.numa_node = memfd->numa;
-      alloc.flags |= CLIB_MEM_VM_F_NUMA_PREFER;
+      clib_unix_warning (0, "memfd ftruncate failure");
+      return SSVM_API_ERROR_CREATE_FAILURE;
     }
-  if ((err = clib_mem_vm_ext_alloc (&alloc)))
+
+  sh = clib_mem_vm_map_shared (uword_to_pointer (memfd->requested_va, void *),
+			       memfd->ssvm_size, memfd->fd, 0,
+			       (char *) memfd->name);
+  if (sh == CLIB_MEM_VM_MAP_FAILED)
     {
-      clib_error_report (err);
+      clib_unix_warning ("memfd map (fd %d)", memfd->fd);
+      close (memfd->fd);
       return SSVM_API_ERROR_CREATE_FAILURE;
     }
 
-  memfd->fd = alloc.fd;
-  memfd->sh = (ssvm_shared_header_t *) alloc.addr;
+  memfd->sh = sh;
   memfd->my_pid = getpid ();
   memfd->i_am_master = 1;
 
-  page_size = 1ull << alloc.log2_page_size;
-  sh = memfd->sh;
   sh->master_pid = memfd->my_pid;
   sh->ssvm_size = memfd->ssvm_size;
   sh->ssvm_va = pointer_to_uword (sh);
   sh->type = SSVM_SEGMENT_MEMFD;
 
+  page_size = 1 << log2_page_size;
   sh->heap = clib_mem_create_heap (((u8 *) sh) + page_size,
 				   memfd->ssvm_size - page_size,
 				   1 /* locked */ , "ssvm master memfd");
@@ -327,7 +344,10 @@ void
 ssvm_delete_memfd (ssvm_private_t * memfd)
 {
   vec_free (memfd->name);
-  clib_mem_vm_free (memfd->sh, memfd->ssvm_size);
+  if (memfd->i_am_master)
+    clib_mem_vm_unmap (memfd->sh);
+  else
+    clib_mem_vm_free (memfd->sh, memfd->ssvm_size);
   close (memfd->fd);
 }
 
@@ -337,34 +357,35 @@ ssvm_delete_memfd (ssvm_private_t * memfd)
 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 };
+  uword page_size, log2_page_size, rnd_size = 0;
   ssvm_shared_header_t *sh;
-  clib_error_t *err;
+  void *oldheap;
   u8 *heap;
 
-  rnd_size = clib_max (ssvm->ssvm_size + (pagesize - 1), ssvm->ssvm_size);
-  rnd_size &= ~(pagesize - 1);
-
-  alloc.name = (char *) ssvm->name;
-  alloc.size = rnd_size + pagesize;
-  if (ssvm->numa)
+  log2_page_size = clib_mem_get_log2_page_size ();
+  if (log2_page_size == 0)
     {
-      alloc.numa_node = ssvm->numa;
-      alloc.flags |= CLIB_MEM_VM_F_NUMA_PREFER;
+      clib_unix_warning (0, "cannot determine page size");
+      return SSVM_API_ERROR_CREATE_FAILURE;
     }
 
-  if ((err = clib_mem_vm_ext_alloc (&alloc)))
+  page_size = 1 << log2_page_size;
+  rnd_size = clib_max (ssvm->ssvm_size + (page_size - 1), ssvm->ssvm_size);
+  rnd_size &= ~(page_size - 1);
+
+  sh = clib_mem_vm_map (0, rnd_size + page_size, log2_page_size,
+			(char *) ssvm->name);
+  if (sh == CLIB_MEM_VM_MAP_FAILED)
     {
-      clib_error_report (err);
+      clib_unix_warning ("private map failed");
       return SSVM_API_ERROR_CREATE_FAILURE;
     }
 
-  heap = clib_mem_create_heap ((u8 *) alloc.addr + pagesize, rnd_size,
+  heap = clib_mem_create_heap ((u8 *) sh + page_size, rnd_size,
 			       1 /* locked */ , "ssvm master private");
   if (heap == 0)
     {
-      clib_unix_warning ("mheap alloc");
+      clib_unix_warning ("heap alloc");
       return -1;
     }
 
@@ -376,7 +397,6 @@ ssvm_master_init_private (ssvm_private_t * ssvm)
   ssvm->requested_va = ~0;
 
   /* First page in allocated memory is set aside for the shared header */
-  sh = alloc.addr;
   ssvm->sh = sh;
 
   clib_memset (sh, 0, sizeof (*sh));
@@ -384,7 +404,10 @@ ssvm_master_init_private (ssvm_private_t * ssvm)
   sh->ssvm_size = rnd_size;
   sh->ssvm_va = pointer_to_uword (heap);
   sh->type = SSVM_SEGMENT_PRIVATE;
-  sh->name = ssvm->name;
+
+  oldheap = ssvm_push_heap (sh);
+  sh->name = format (0, "%s", ssvm->name, 0);
+  ssvm_pop_heap (oldheap);
 
   return 0;
 }
@@ -401,7 +424,7 @@ ssvm_delete_private (ssvm_private_t * ssvm)
 {
   vec_free (ssvm->name);
   clib_mem_destroy_heap (ssvm->sh->heap);
-  clib_mem_vm_free (ssvm->sh, ssvm->ssvm_size + clib_mem_get_page_size ());
+  clib_mem_vm_unmap (ssvm->sh);
 }
 
 int
diff --git a/src/svm/ssvm.h b/src/svm/ssvm.h
index 2a9da3d2f32..3a277cb5c5f 100644
--- a/src/svm/ssvm.h
+++ b/src/svm/ssvm.h
@@ -85,7 +85,7 @@ typedef struct
   uword requested_va;
   u32 my_pid;
   u8 *name;
-  u8 numa;			/**< Numa requested at alloc time */
+  u8 numa;			/**< UNUSED: numa requested at alloc time */
   int i_am_master;
 
   union
diff --git a/src/vnet/session/application.c b/src/vnet/session/application.c
index a0ddaed90cf..0b5c951fd70 100644
--- a/src/vnet/session/application.c
+++ b/src/vnet/session/application.c
@@ -876,11 +876,8 @@ vnet_application_attach (vnet_app_attach_args_t * a)
   if (application_is_proxy (app))
     {
       application_setup_proxy (app);
-      /*
-       * I suspect the segment manager pool gets reallocated (because a new listener
-       * is added, and each listener has a segment manager) so sm becomes a dangling
-       * reference. That's why we need to re-grab sm.
-       */
+      /* The segment manager pool is reallocated because a new listener
+       * is added. Re-grab segment manager to avoid dangling reference */
       sm = segment_manager_get (app_wrk->first_segment_manager);
     }
 
diff --git a/src/vnet/session/application_worker.c b/src/vnet/session/application_worker.c
index 17aeb57d49b..2b2cc344f05 100644
--- a/src/vnet/session/application_worker.c
+++ b/src/vnet/session/application_worker.c
@@ -180,12 +180,12 @@ app_worker_alloc_segment_manager (app_worker_t * app_wrk)
     {
       sm = segment_manager_get (app_wrk->first_segment_manager);
       app_wrk->first_segment_manager_in_use = 1;
-      return sm;
     }
-
-  sm = segment_manager_alloc ();
+  else
+    {
+      sm = segment_manager_alloc ();
+    }
   sm->app_wrk_index = app_wrk->wrk_index;
-
   return sm;
 }
 
diff --git a/src/vnet/session/segment_manager.c b/src/vnet/session/segment_manager.c
index dbeff6e4acf..44ff5fefd70 100644
--- a/src/vnet/session/segment_manager.c
+++ b/src/vnet/session/segment_manager.c
@@ -138,11 +138,16 @@ segment_manager_add_segment (segment_manager_t * sm, uword segment_size)
 	}
     }
   else
-    seg_name = format (0, "%s%c", "process-private", 0);
+    {
+      app_worker_t *app_wrk = app_worker_get (sm->app_wrk_index);
+      application_t *app = application_get (app_wrk->app_index);
+      seg_name = format (0, "%v segment%c", app->name, 0);
+    }
 
   fs->ssvm.ssvm_size = segment_size;
   fs->ssvm.name = seg_name;
-  fs->ssvm.requested_va = baseva;
+  /* clib_mem_vm_map_shared consumes first page before requested_va */
+  fs->ssvm.requested_va = baseva + page_size;
 
   if ((rv = ssvm_master_init (&fs->ssvm, props->segment_type)))
     {
@@ -192,7 +197,9 @@ segment_manager_del_segment (segment_manager_t * sm, fifo_segment_t * fs)
 
   if (ssvm_type (&fs->ssvm) != SSVM_SEGMENT_PRIVATE)
     {
-      clib_valloc_free (&smm->va_allocator, fs->ssvm.requested_va);
+      /* clib_mem_vm_map_shared consumes first page before requested_va */
+      clib_valloc_free (&smm->va_allocator,
+			fs->ssvm.requested_va - clib_mem_get_page_size ());
 
       if (!segment_manager_app_detached (sm))
 	{
@@ -947,6 +954,7 @@ segment_manager_show_fn (vlib_main_t * vm, unformat_input_t * input,
       }));
       /* *INDENT-ON* */
 
+      vlib_cli_output (vm, "\n");
     }
   if (show_segments)
     {
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index d27d1bbab23..5065f7834b1 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -1496,8 +1496,9 @@ session_vpp_event_queues_allocate (session_main_t * smm)
       eqs->ssvm_size = eqs_size;
       eqs->i_am_master = 1;
       eqs->my_pid = vpp_pid;
-      eqs->name = format (0, "%s%c", "evt-qs-segment", 0);
-      eqs->requested_va = smm->session_baseva;
+      eqs->name = format (0, "%s%c", "session: evt-qs-segment", 0);
+      /* clib_mem_vm_map_shared consumes first page before requested_va */
+      eqs->requested_va = smm->session_baseva + clib_mem_get_page_size ();
 
       if (ssvm_master_init (eqs, SSVM_SEGMENT_MEMFD))
 	{
@@ -1661,6 +1662,7 @@ session_manager_main_enable (vlib_main_t * vm)
   session_main_t *smm = &session_main;
   vlib_thread_main_t *vtm = vlib_get_thread_main ();
   u32 num_threads, preallocated_sessions_per_worker;
+  uword margin = 8 << 12;
   session_worker_t *wrk;
   int i;
 
@@ -1691,7 +1693,7 @@ session_manager_main_enable (vlib_main_t * vm)
   session_vpp_event_queues_allocate (smm);
 
   /* Initialize fifo segment main baseva and timeout */
-  sm_args->baseva = smm->session_baseva + smm->evt_qs_segment_size;
+  sm_args->baseva = smm->session_baseva + smm->evt_qs_segment_size + margin;
   sm_args->size = smm->session_va_space_size;
   segment_manager_main_init (sm_args);