VPP-42: VPP crashes in IPsec code when running multithreaded

Change-Id: Ib231642cfead5f5e8e45508361a11c87aad83b51
Signed-off-by: Matthew Smith <mgsmith@netgate.com>

5 files changed, 58 insertions, 33 deletions

diff --git a/vnet/vnet/ipsec/esp.h b/vnet/vnet/ipsec/esp.h
index b44c26e1..2334dc48 100644
--- a/vnet/vnet/ipsec/esp.h
+++ b/vnet/vnet/ipsec/esp.h
@@ -50,16 +50,22 @@ typedef struct {
   u8 trunc_size;
 } esp_integ_alg_t;
 
-
 typedef struct {
-  esp_crypto_alg_t * esp_crypto_algs;
-  esp_integ_alg_t * esp_integ_algs;
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline0);
   EVP_CIPHER_CTX encrypt_ctx;
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline1);
   EVP_CIPHER_CTX decrypt_ctx;
+  CLIB_CACHE_LINE_ALIGN_MARK(cacheline2);
   HMAC_CTX hmac_ctx;
-  ipsec_crypto_alg_t last_encrytp_alg;
-  ipsec_crypto_alg_t last_decrytp_alg;
+  ipsec_crypto_alg_t last_encrypt_alg;
+  ipsec_crypto_alg_t last_decrypt_alg;
   ipsec_integ_alg_t last_integ_alg;
+} esp_main_per_thread_data_t;
+
+typedef struct {
+  esp_crypto_alg_t * esp_crypto_algs;
+  esp_integ_alg_t * esp_integ_algs;
+  esp_main_per_thread_data_t * per_thread_data;
 } esp_main_t;
 
 esp_main_t esp_main;
@@ -68,6 +74,7 @@ always_inline void
 esp_init()
 {
   esp_main_t * em = &esp_main;
+  vlib_thread_main_t * tm = vlib_get_thread_main();
 
   memset (em, 0, sizeof (em[0]));
 
@@ -99,9 +106,15 @@ esp_init()
   i->md = EVP_sha512();
   i->trunc_size = 32;
 
-  EVP_CIPHER_CTX_init(&(em->encrypt_ctx));
-  EVP_CIPHER_CTX_init(&(em->decrypt_ctx));
-  HMAC_CTX_init(&(em->hmac_ctx));
+  vec_validate_aligned(em->per_thread_data, tm->n_vlib_mains-1, CLIB_CACHE_LINE_BYTES);
+  int thread_id;
+
+  for (thread_id = 0; thread_id < tm->n_vlib_mains - 1; thread_id++)
+    {
+      EVP_CIPHER_CTX_init(&(em->per_thread_data[thread_id].encrypt_ctx));
+      EVP_CIPHER_CTX_init(&(em->per_thread_data[thread_id].decrypt_ctx));
+      HMAC_CTX_init(&(em->per_thread_data[thread_id].hmac_ctx));
+    }
 }
 
 always_inline unsigned int
@@ -115,7 +128,8 @@ hmac_calc(ipsec_integ_alg_t alg,
           u32 seq_hi)
 {
   esp_main_t * em = &esp_main;
-  HMAC_CTX * ctx = &(em->hmac_ctx);
+  u32 cpu_index = os_get_cpu_number();
+  HMAC_CTX * ctx = &(em->per_thread_data[cpu_index].hmac_ctx);
   const EVP_MD * md = NULL;
   unsigned int len;
 
@@ -124,9 +138,9 @@ hmac_calc(ipsec_integ_alg_t alg,
   if (PREDICT_FALSE(em->esp_integ_algs[alg].md == 0))
     return 0;
 
-  if (PREDICT_FALSE(alg != em->last_integ_alg)) {
+  if (PREDICT_FALSE(alg != em->per_thread_data[cpu_index].last_integ_alg)) {
     md = em->esp_integ_algs[alg].md;
-    em->last_integ_alg = alg;
+    em->per_thread_data[cpu_index].last_integ_alg = alg;
   }
 
   HMAC_Init(ctx, key, key_len, md);
diff --git a/vnet/vnet/ipsec/esp_decrypt.c b/vnet/vnet/ipsec/esp_decrypt.c
index ad511b0f..958a4d67 100644
--- a/vnet/vnet/ipsec/esp_decrypt.c
+++ b/vnet/vnet/ipsec/esp_decrypt.c
@@ -85,7 +85,8 @@ esp_decrypt_aes_cbc(ipsec_crypto_alg_t alg,
                     u8 * iv)
 {
   esp_main_t * em = &esp_main;
-  EVP_CIPHER_CTX * ctx = &(em->decrypt_ctx);
+  u32 cpu_index = os_get_cpu_number();
+  EVP_CIPHER_CTX * ctx = &(em->per_thread_data[cpu_index].decrypt_ctx);
   const EVP_CIPHER * cipher = NULL;
   int out_len;
 
@@ -94,9 +95,9 @@ esp_decrypt_aes_cbc(ipsec_crypto_alg_t alg,
   if (PREDICT_FALSE(em->esp_crypto_algs[alg].type == 0))
     return;
 
-  if (PREDICT_FALSE(alg != em->last_decrytp_alg)) {
+  if (PREDICT_FALSE(alg != em->per_thread_data[cpu_index].last_decrypt_alg)) {
     cipher = em->esp_crypto_algs[alg].type;
-    em->last_decrytp_alg = alg;
+    em->per_thread_data[cpu_index].last_decrypt_alg = alg;
   }
 
   EVP_DecryptInit_ex(ctx, cipher, NULL, key, iv);
@@ -235,10 +236,12 @@ esp_decrypt_node_fn (vlib_main_t * vm,
   u32 * recycle = 0;
   from = vlib_frame_vector_args (from_frame);
   n_left_from = from_frame->n_vectors;
+  u32 cpu_index = os_get_cpu_number();
+  u32 * empty_buffers = im->empty_buffers[cpu_index];
 
   ipsec_alloc_empty_buffers(vm, im);
 
-  if (PREDICT_FALSE(vec_len (im->empty_buffers) < n_left_from)){
+  if (PREDICT_FALSE(vec_len (empty_buffers) < n_left_from)){
     vlib_node_increment_counter (vm, esp_decrypt_node.index,
                                  ESP_DECRYPT_ERROR_NO_BUFFER, n_left_from);
     goto free_buffers_and_exit;
@@ -327,11 +330,11 @@ esp_decrypt_node_fn (vlib_main_t * vm,
              }
 
           /* grab free buffer */
-          uword last_empty_buffer = vec_len (im->empty_buffers) - 1;
-          o_bi0 = im->empty_buffers[last_empty_buffer];
+          uword last_empty_buffer = vec_len (empty_buffers) - 1;
+          o_bi0 = empty_buffers[last_empty_buffer];
           o_b0 = vlib_get_buffer (vm, o_bi0);
-          vlib_prefetch_buffer_with_index (vm, im->empty_buffers[last_empty_buffer-1], STORE);
-          _vec_len (im->empty_buffers) = last_empty_buffer;
+          vlib_prefetch_buffer_with_index (vm, empty_buffers[last_empty_buffer-1], STORE);
+          _vec_len (empty_buffers) = last_empty_buffer;
 
           /* add old buffer to the recycle list */
           vec_add1(recycle, i_bi0);
diff --git a/vnet/vnet/ipsec/esp_encrypt.c b/vnet/vnet/ipsec/esp_encrypt.c
index 7194399f..39bbf2e4 100644
--- a/vnet/vnet/ipsec/esp_encrypt.c
+++ b/vnet/vnet/ipsec/esp_encrypt.c
@@ -89,7 +89,8 @@ esp_encrypt_aes_cbc(ipsec_crypto_alg_t alg,
                     u8 * iv)
 {
   esp_main_t * em = &esp_main;
-  EVP_CIPHER_CTX * ctx = &(em->encrypt_ctx);
+  u32 cpu_index = os_get_cpu_number();
+  EVP_CIPHER_CTX * ctx = &(em->per_thread_data[cpu_index].encrypt_ctx);
   const EVP_CIPHER * cipher = NULL;
   int out_len;
 
@@ -98,9 +99,9 @@ esp_encrypt_aes_cbc(ipsec_crypto_alg_t alg,
   if (PREDICT_FALSE(em->esp_crypto_algs[alg].type == IPSEC_CRYPTO_ALG_NONE))
     return;
 
-  if (PREDICT_FALSE(alg != em->last_encrytp_alg)) {
+  if (PREDICT_FALSE(alg != em->per_thread_data[cpu_index].last_encrypt_alg)) {
     cipher = em->esp_crypto_algs[alg].type;
-    em->last_encrytp_alg = alg;
+    em->per_thread_data[cpu_index].last_encrypt_alg = alg;
   }
 
   EVP_EncryptInit_ex(ctx, cipher, NULL, key, iv);
@@ -142,10 +143,12 @@ esp_encrypt_node_fn (vlib_main_t * vm,
   n_left_from = from_frame->n_vectors;
   ipsec_main_t *im = &ipsec_main;
   u32 * recycle = 0;
+  u32 cpu_index = os_get_cpu_number();
+  u32 * empty_buffers = im->empty_buffers[cpu_index];
 
   ipsec_alloc_empty_buffers(vm, im);
 
-  if (PREDICT_FALSE(vec_len (im->empty_buffers) < n_left_from)){
+  if (PREDICT_FALSE(vec_len (empty_buffers) < n_left_from)){
     vlib_node_increment_counter (vm, esp_encrypt_node.index,
                                  ESP_ENCRYPT_ERROR_NO_BUFFER, n_left_from);
     clib_warning("no enough empty buffers. discarding frame");
@@ -197,13 +200,13 @@ esp_encrypt_node_fn (vlib_main_t * vm,
             }
 
           /* grab free buffer */
-          last_empty_buffer = vec_len (im->empty_buffers) - 1;
-          o_bi0 = im->empty_buffers[last_empty_buffer];
+          last_empty_buffer = vec_len (empty_buffers) - 1;
+          o_bi0 = empty_buffers[last_empty_buffer];
           o_b0 = vlib_get_buffer (vm, o_bi0);
           o_b0->current_data = sizeof(ethernet_header_t);
           ih0 = vlib_buffer_get_current (i_b0);
-          vlib_prefetch_buffer_with_index (vm, im->empty_buffers[last_empty_buffer-1], STORE);
-          _vec_len (im->empty_buffers) = last_empty_buffer;
+          vlib_prefetch_buffer_with_index (vm, empty_buffers[last_empty_buffer-1], STORE);
+          _vec_len (empty_buffers) = last_empty_buffer;
           to_next[0] = o_bi0;
           to_next += 1;
 
diff --git a/vnet/vnet/ipsec/ipsec.c b/vnet/vnet/ipsec/ipsec.c
index 47beafc3..ea077d0a 100644
--- a/vnet/vnet/ipsec/ipsec.c
+++ b/vnet/vnet/ipsec/ipsec.c
@@ -492,6 +492,7 @@ ipsec_init (vlib_main_t * vm)
 {
   clib_error_t * error;
   ipsec_main_t * im = &ipsec_main;
+  vlib_thread_main_t * tm = vlib_get_thread_main();
   vlib_node_t * node;
 
   ipsec_rand_seed();
@@ -505,6 +506,8 @@ ipsec_init (vlib_main_t * vm)
   im->sa_index_by_sa_id        = hash_create (0, sizeof (uword));
   im->spd_index_by_sw_if_index = hash_create (0, sizeof (uword));
 
+  vec_validate_aligned(im->empty_buffers, tm->n_vlib_mains-1, CLIB_CACHE_LINE_BYTES);
+
   node = vlib_get_node_by_name (vm, (u8 *) "error-drop");
   ASSERT(node);
   im->error_drop_node_index = node->index;
diff --git a/vnet/vnet/ipsec/ipsec.h b/vnet/vnet/ipsec/ipsec.h
index 386c0ade..8cd2b41c 100644
--- a/vnet/vnet/ipsec/ipsec.h
+++ b/vnet/vnet/ipsec/ipsec.h
@@ -175,7 +175,7 @@ typedef struct {
   ipsec_tunnel_if_t * tunnel_interfaces;
   u32 * free_tunnel_if_indices;
 
-  u32 * empty_buffers;
+  u32 ** empty_buffers;
 
   uword * tunnel_index_by_key;
 
@@ -242,20 +242,22 @@ ipsec_alloc_empty_buffers(vlib_main_t * vm, ipsec_main_t *im)
 #else
   u32 free_list_index = VLIB_BUFFER_DEFAULT_FREE_LIST_INDEX;
 #endif
-  uword l = vec_len (im->empty_buffers);
+  u32 cpu_index = os_get_cpu_number();
+  uword l = vec_len (im->empty_buffers[cpu_index]);
   uword n_alloc = 0;
 
   if (PREDICT_FALSE(l < VLIB_FRAME_SIZE))
     {
-      if (!im->empty_buffers) {
-        vec_alloc (im->empty_buffers, 2 * VLIB_FRAME_SIZE );
+      if (!im->empty_buffers[cpu_index]) {
+        vec_alloc (im->empty_buffers[cpu_index], 2 * VLIB_FRAME_SIZE );
       }
 
-      n_alloc = vlib_buffer_alloc_from_free_list (vm, im->empty_buffers + l,
+      n_alloc = vlib_buffer_alloc_from_free_list (vm,
+                                                  im->empty_buffers[cpu_index] + l,
                                                   2 * VLIB_FRAME_SIZE - l,
                                                   free_list_index);
 
-      _vec_len (im->empty_buffers) = l + n_alloc;
+      _vec_len (im->empty_buffers[cpu_index]) = l + n_alloc;
     }
 }