diff options
Diffstat (limited to 'src/vlibmemory')
| -rw-r--r-- | src/vlibmemory/api.h | 163 | ||||
| -rw-r--r-- | src/vlibmemory/memclnt.api | 91 | ||||
| -rw-r--r-- | src/vlibmemory/memory_client.c | 283 | ||||
| -rw-r--r-- | src/vlibmemory/memory_shared.c | 852 | ||||
| -rw-r--r-- | src/vlibmemory/memory_vlib.c | 1346 | ||||
| -rw-r--r-- | src/vlibmemory/unix_shared_memory_queue.c | 324 | ||||
| -rw-r--r-- | src/vlibmemory/unix_shared_memory_queue.h | 69 | ||||
| -rw-r--r-- | src/vlibmemory/vl_memory_api_h.h | 32 | ||||
| -rw-r--r-- | src/vlibmemory/vl_memory_msg_enum.h | 42 |
9 files changed, 3202 insertions, 0 deletions
diff --git a/src/vlibmemory/api.h b/src/vlibmemory/api.h new file mode 100644 index 00000000000..54a0a0011fd --- /dev/null +++ b/src/vlibmemory/api.h @@ -0,0 +1,163 @@ +/* + *------------------------------------------------------------------ + * api.h + * + * Copyright (c) 2009 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#ifndef included_vlibmemory_api_h +#define included_vlibmemory_api_h + +#include <vppinfra/error.h> +#include <svm/svm.h> +#include <vlib/vlib.h> +#include <vlibmemory/unix_shared_memory_queue.h> +#include <vlib/unix/unix.h> +#include <vlibapi/api.h> + +/* Allocated in shared memory */ + +/* + * Ring-allocation scheme for client API messages + * + * Only one proc/thread has control of a given message buffer. + * To free a buffer allocated from one of these rings, we clear + * a field in the buffer (header), and leave. + * + * No locks, no hits, no errors... + */ +typedef struct ring_alloc_ +{ + unix_shared_memory_queue_t *rp; + u16 size; + u16 nitems; + u32 hits; + u32 misses; +} ring_alloc_t; + +/* + * Initializers for the (shared-memory) rings + * _(size, n). Note: each msg has an 8 byte header. + * Might want to change that to an index sometime. + */ +#define foreach_vl_aring_size \ +_(64+8, 1024) \ +_(256+8, 128) \ +_(1024+8, 64) + +#define foreach_clnt_aring_size \ +_(1024+8, 1024) \ +_(2048+8, 128) \ +_(4096+8, 8) + +typedef struct vl_shmem_hdr_ +{ + int version; + + /* getpid () for the VLIB client process */ + volatile int vl_pid; + + /* Client sends VLIB msgs here. */ + unix_shared_memory_queue_t *vl_input_queue; + + /* Vector of rings; one for each size. */ + + /* VLIB allocates buffers to send msgs to clients here. */ + ring_alloc_t *vl_rings; + + /* Clients allocate buffer to send msgs to VLIB here. */ + ring_alloc_t *client_rings; + + /* Number of detected application restarts */ + u32 application_restarts; + + /* Number of messages reclaimed during application restart */ + u32 restart_reclaims; + +} vl_shmem_hdr_t; + +#define VL_SHM_VERSION 2 + +#define VL_API_EPOCH_MASK 0xFF +#define VL_API_EPOCH_SHIFT 8 + +static inline u32 +vl_msg_api_handle_get_epoch (u32 index) +{ + return (index & VL_API_EPOCH_MASK); +} + +static inline u32 +vl_msg_api_handle_get_index (u32 index) +{ + return (index >> VL_API_EPOCH_SHIFT); +} + +static inline u32 +vl_msg_api_handle_from_index_and_epoch (u32 index, u32 epoch) +{ + u32 handle; + ASSERT (index < 0x00FFFFFF); + + handle = (index << VL_API_EPOCH_SHIFT) | (epoch & VL_API_EPOCH_MASK); + return handle; +} + +void *vl_msg_api_alloc (int nbytes); +void *vl_msg_api_alloc_or_null (int nbytes); +void *vl_msg_api_alloc_as_if_client (int nbytes); +void *vl_msg_api_alloc_as_if_client_or_null (int nbytes); +void vl_msg_api_free (void *a); +int vl_map_shmem (char *region_name, int is_vlib); +void vl_register_mapped_shmem_region (svm_region_t * rp); +void vl_unmap_shmem (void); +void vl_msg_api_send_shmem (unix_shared_memory_queue_t * q, u8 * elem); +void vl_msg_api_send_shmem_nolock (unix_shared_memory_queue_t * q, u8 * elem); +void vl_msg_api_send (vl_api_registration_t * rp, u8 * elem); +int vl_client_connect (char *name, int ctx_quota, int input_queue_size); +void vl_client_disconnect (void); +unix_shared_memory_queue_t *vl_api_client_index_to_input_queue (u32 index); +vl_api_registration_t *vl_api_client_index_to_registration (u32 index); +int vl_client_api_map (char *region_name); +void vl_client_api_unmap (void); +void vl_set_memory_region_name (char *name); +void vl_set_memory_root_path (char *root_path); +void vl_set_memory_uid (int uid); +void vl_set_memory_gid (int gid); +void vl_set_global_memory_baseva (u64 baseva); +void vl_set_global_memory_size (u64 size); +void vl_set_api_memory_size (u64 size); +void vl_set_global_pvt_heap_size (u64 size); +void vl_set_api_pvt_heap_size (u64 size); +void vl_enable_disable_memory_api (vlib_main_t * vm, int yesno); +void vl_client_disconnect_from_vlib (void); +int vl_client_connect_to_vlib (char *svm_name, char *client_name, + int rx_queue_size); +int vl_client_connect_to_vlib_no_rx_pthread (char *svm_name, + char *client_name, + int rx_queue_size); +u16 vl_client_get_first_plugin_msg_id (char *plugin_name); + +void vl_api_rpc_call_main_thread (void *fp, u8 * data, u32 data_length); + +#endif /* included_vlibmemory_api_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/memclnt.api b/src/vlibmemory/memclnt.api new file mode 100644 index 00000000000..2f654caf2c1 --- /dev/null +++ b/src/vlibmemory/memclnt.api @@ -0,0 +1,91 @@ +/* Hey Emacs use -*- mode: C -*- */ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * Create a client registration + */ +manual_print +define memclnt_create { + i32 ctx_quota; /* requested punt context quota */ + u32 context; /* opaque value to be returned in the reply */ + u64 input_queue; /* client's queue */ + u8 name[64]; /* for show, find by name, whatever */ + u32 api_versions[8]; /* client-server pairs use as desired */ +}; + +define memclnt_create_reply { + i32 response; /* Non-negative = success */ + u64 handle; /* handle by which vlib knows this client */ + u32 index; /* index, used e.g. by API trace replay */ + u32 context; /* opaque value from the create request */ + u64 message_table; /* serialized message table in shmem */ +}; + +/* + * Delete a client registration + */ +manual_print +define memclnt_delete { + u32 index; /* index, used e.g. by API trace replay */ + u64 handle; /* handle by which vlib knows this client */ +}; + +define memclnt_delete_reply { + i32 response; /* Non-negative = success */ + u64 handle; /* in case the client wonders */ +}; + +/* + * Client RX thread exit + */ + +define rx_thread_exit { + u8 dummy; +}; + +/* + * RPC + */ +define rpc_call { + u32 client_index; + u32 context; + u64 function; + u8 multicast; + u8 need_barrier_sync; + u8 send_reply; + u8 data[0]; +}; + +define rpc_reply { + i32 retval; + u32 context; +}; + +/* + * Lookup message-ID base by name + */ +define get_first_msg_id { + u32 client_index; + u32 context; + u8 name[64]; +}; + +define get_first_msg_id_reply { + u32 client_index; + u32 context; + i32 retval; + u16 first_msg_id; +}; diff --git a/src/vlibmemory/memory_client.c b/src/vlibmemory/memory_client.c new file mode 100644 index 00000000000..234a0a5a330 --- /dev/null +++ b/src/vlibmemory/memory_client.c @@ -0,0 +1,283 @@ +/* + *------------------------------------------------------------------ + * memory_client.c - API message handling, client code. + * + * Copyright (c) 2010 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include <stdio.h> +#include <stdlib.h> +#include <setjmp.h> +#include <sys/types.h> +#include <sys/mman.h> +#include <sys/stat.h> +#include <netinet/in.h> +#include <signal.h> +#include <pthread.h> +#include <unistd.h> +#include <time.h> +#include <fcntl.h> +#include <string.h> +#include <vppinfra/clib.h> +#include <vppinfra/vec.h> +#include <vppinfra/hash.h> +#include <vppinfra/bitmap.h> +#include <vppinfra/fifo.h> +#include <vppinfra/time.h> +#include <vppinfra/mheap.h> +#include <vppinfra/heap.h> +#include <vppinfra/pool.h> +#include <vppinfra/format.h> + +#include <vlib/vlib.h> +#include <vlib/unix/unix.h> +#include <vlibmemory/api.h> + +#include <vlibmemory/vl_memory_msg_enum.h> + +#define vl_typedefs /* define message structures */ +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_typedefs + +#define vl_endianfun /* define message structures */ +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_endianfun + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) clib_warning (__VA_ARGS__) +#define vl_printfun +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_printfun + +typedef struct +{ + u8 rx_thread_jmpbuf_valid; + u8 connected_to_vlib; + jmp_buf rx_thread_jmpbuf; + pthread_t rx_thread_handle; + /* Plugin message base lookup scheme */ + volatile u8 first_msg_id_reply_ready; + u16 first_msg_id_reply; +} memory_client_main_t; + +memory_client_main_t memory_client_main; + +static void * +rx_thread_fn (void *arg) +{ + unix_shared_memory_queue_t *q; + memory_client_main_t *mm = &memory_client_main; + api_main_t *am = &api_main; + + q = am->vl_input_queue; + + /* So we can make the rx thread terminate cleanly */ + if (setjmp (mm->rx_thread_jmpbuf) == 0) + { + mm->rx_thread_jmpbuf_valid = 1; + while (1) + { + vl_msg_api_queue_handler (q); + } + } + pthread_exit (0); +} + +static void +vl_api_rx_thread_exit_t_handler (vl_api_rx_thread_exit_t * mp) +{ + memory_client_main_t *mm = &memory_client_main; + vl_msg_api_free (mp); + longjmp (mm->rx_thread_jmpbuf, 1); +} + +static void +noop_handler (void *notused) +{ +} + +#define foreach_api_msg \ +_(RX_THREAD_EXIT, rx_thread_exit) + +static int +connect_to_vlib_internal (char *svm_name, char *client_name, + int rx_queue_size, int want_pthread) +{ + int rv = 0; + memory_client_main_t *mm = &memory_client_main; + + if ((rv = vl_client_api_map (svm_name))) + { + clib_warning ("vl_client_api map rv %d", rv); + return rv; + } + +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, #n, \ + vl_api_##n##_t_handler, \ + noop_handler, \ + vl_api_##n##_t_endian, \ + vl_api_##n##_t_print, \ + sizeof(vl_api_##n##_t), 1); + foreach_api_msg; +#undef _ + + if (vl_client_connect (client_name, 0 /* punt quota */ , + rx_queue_size /* input queue */ ) < 0) + { + vl_client_api_unmap (); + return -1; + } + + /* Start the rx queue thread */ + + if (want_pthread) + { + rv = pthread_create (&mm->rx_thread_handle, + NULL /*attr */ , rx_thread_fn, 0); + if (rv) + clib_warning ("pthread_create returned %d", rv); + } + + mm->connected_to_vlib = 1; + return 0; +} + +int +vl_client_connect_to_vlib (char *svm_name, char *client_name, + int rx_queue_size) +{ + return connect_to_vlib_internal (svm_name, client_name, rx_queue_size, + 1 /* want pthread */ ); +} + +int +vl_client_connect_to_vlib_no_rx_pthread (char *svm_name, char *client_name, + int rx_queue_size) +{ + return connect_to_vlib_internal (svm_name, client_name, rx_queue_size, + 0 /* want pthread */ ); +} + +void +vl_client_disconnect_from_vlib (void) +{ + memory_client_main_t *mm = &memory_client_main; + api_main_t *am = &api_main; + uword junk; + + if (mm->rx_thread_jmpbuf_valid) + { + vl_api_rx_thread_exit_t *ep; + ep = vl_msg_api_alloc (sizeof (*ep)); + ep->_vl_msg_id = ntohs (VL_API_RX_THREAD_EXIT); + vl_msg_api_send_shmem (am->vl_input_queue, (u8 *) & ep); + pthread_join (mm->rx_thread_handle, (void **) &junk); + } + if (mm->connected_to_vlib) + { + vl_client_disconnect (); + vl_client_api_unmap (); + } + memset (mm, 0, sizeof (*mm)); +} + +static void vl_api_get_first_msg_id_reply_t_handler + (vl_api_get_first_msg_id_reply_t * mp) +{ + memory_client_main_t *mm = &memory_client_main; + i32 retval = ntohl (mp->retval); + + mm->first_msg_id_reply = (retval >= 0) ? ntohs (mp->first_msg_id) : ~0; + mm->first_msg_id_reply_ready = 1; +} + +u16 +vl_client_get_first_plugin_msg_id (char *plugin_name) +{ + vl_api_get_first_msg_id_t *mp; + api_main_t *am = &api_main; + memory_client_main_t *mm = &memory_client_main; + f64 timeout; + void *old_handler; + clib_time_t clib_time; + u16 rv = ~0; + + if (strlen (plugin_name) + 1 > sizeof (mp->name)) + return (rv); + + memset (&clib_time, 0, sizeof (clib_time)); + clib_time_init (&clib_time); + + /* Push this plugin's first_msg_id_reply handler */ + old_handler = am->msg_handlers[VL_API_GET_FIRST_MSG_ID_REPLY]; + am->msg_handlers[VL_API_GET_FIRST_MSG_ID_REPLY] = (void *) + vl_api_get_first_msg_id_reply_t_handler; + + /* Ask the data-plane for the message-ID base of the indicated plugin */ + mm->first_msg_id_reply_ready = 0; + + mp = vl_msg_api_alloc (sizeof (*mp)); + memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = ntohs (VL_API_GET_FIRST_MSG_ID); + mp->client_index = am->my_client_index; + strncpy ((char *) mp->name, plugin_name, sizeof (mp->name) - 1); + + vl_msg_api_send_shmem (am->shmem_hdr->vl_input_queue, (u8 *) & mp); + + /* Synchronously wait for the answer */ + do + { + timeout = clib_time_now (&clib_time) + 1.0; + + while (clib_time_now (&clib_time) < timeout) + { + if (mm->first_msg_id_reply_ready == 1) + { + rv = mm->first_msg_id_reply; + goto result; + } + } + /* Restore old handler */ + am->msg_handlers[VL_API_GET_FIRST_MSG_ID_REPLY] = old_handler; + + return rv; + } + while (0); + +result: + + /* Restore the old handler */ + am->msg_handlers[VL_API_GET_FIRST_MSG_ID_REPLY] = old_handler; + + if (rv == (u16) ~ 0) + clib_warning ("plugin '%s' not registered", plugin_name); + + return rv; +} + +void +vlib_node_sync_stats (vlib_main_t * vm, vlib_node_t * n) +{ + clib_warning ("STUB called..."); +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/memory_shared.c b/src/vlibmemory/memory_shared.c new file mode 100644 index 00000000000..d8d3200485e --- /dev/null +++ b/src/vlibmemory/memory_shared.c @@ -0,0 +1,852 @@ +/* + *------------------------------------------------------------------ + * memclnt_shared.c - API message handling, common code for both clients + * and the vlib process itself. + * + * + * Copyright (c) 2009 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include <stdio.h> +#include <stdlib.h> +#include <stddef.h> +#include <string.h> +#include <unistd.h> +#include <signal.h> +#include <vppinfra/format.h> +#include <vppinfra/byte_order.h> +#include <vppinfra/error.h> +#include <vlib/vlib.h> +#include <vlib/unix/unix.h> +#include <vlibmemory/api.h> +#include <vlibmemory/unix_shared_memory_queue.h> + +#include <vlibmemory/vl_memory_msg_enum.h> + +#define vl_typedefs +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_typedefs + +static inline void * +vl_msg_api_alloc_internal (int nbytes, int pool, int may_return_null) +{ + int i; + msgbuf_t *rv; + ring_alloc_t *ap; + unix_shared_memory_queue_t *q; + void *oldheap; + vl_shmem_hdr_t *shmem_hdr; + api_main_t *am = &api_main; + + shmem_hdr = am->shmem_hdr; + + if (shmem_hdr == 0) + { + clib_warning ("shared memory header NULL"); + return 0; + } + + /* account for the msgbuf_t header */ + nbytes += sizeof (msgbuf_t); + + if (shmem_hdr->vl_rings == 0) + { + clib_warning ("vl_rings NULL"); + ASSERT (0); + abort (); + } + + if (shmem_hdr->client_rings == 0) + { + clib_warning ("client_rings NULL"); + ASSERT (0); + abort (); + } + + ap = pool ? shmem_hdr->vl_rings : shmem_hdr->client_rings; + for (i = 0; i < vec_len (ap); i++) + { + /* Too big? */ + if (nbytes > ap[i].size) + { + continue; + } + + q = ap[i].rp; + if (pool == 0) + { + pthread_mutex_lock (&q->mutex); + } + rv = (msgbuf_t *) (&q->data[0] + q->head * q->elsize); + /* + * Is this item still in use? + */ + if (rv->q) + { + /* yes, loser; try next larger pool */ + ap[i].misses++; + if (pool == 0) + pthread_mutex_unlock (&q->mutex); + continue; + } + /* OK, we have a winner */ + ap[i].hits++; + /* + * Remember the source queue, although we + * don't need to know the queue to free the item. + */ + rv->q = q; + q->head++; + if (q->head == q->maxsize) + q->head = 0; + + if (pool == 0) + pthread_mutex_unlock (&q->mutex); + goto out; + } + + /* + * Request too big, or head element of all size-compatible rings + * still in use. Fall back to shared-memory malloc. + */ + am->ring_misses++; + + pthread_mutex_lock (&am->vlib_rp->mutex); + oldheap = svm_push_data_heap (am->vlib_rp); + if (may_return_null) + { + rv = clib_mem_alloc_or_null (nbytes); + if (PREDICT_FALSE (rv == 0)) + { + svm_pop_heap (oldheap); + pthread_mutex_unlock (&am->vlib_rp->mutex); + return 0; + } + } + else + rv = clib_mem_alloc (nbytes); + + rv->q = 0; + svm_pop_heap (oldheap); + pthread_mutex_unlock (&am->vlib_rp->mutex); + +out: + rv->data_len = htonl (nbytes - sizeof (msgbuf_t)); + return (rv->data); +} + +void * +vl_msg_api_alloc (int nbytes) +{ + int pool; + api_main_t *am = &api_main; + vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr; + + /* + * Clients use pool-0, vlib proc uses pool 1 + */ + pool = (am->our_pid == shmem_hdr->vl_pid); + return vl_msg_api_alloc_internal (nbytes, pool, 0 /* may_return_null */ ); +} + +void * +vl_msg_api_alloc_or_null (int nbytes) +{ + int pool; + api_main_t *am = &api_main; + vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr; + + pool = (am->our_pid == shmem_hdr->vl_pid); + return vl_msg_api_alloc_internal (nbytes, pool, 1 /* may_return_null */ ); +} + +void * +vl_msg_api_alloc_as_if_client (int nbytes) +{ + return vl_msg_api_alloc_internal (nbytes, 0, 0 /* may_return_null */ ); +} + +void * +vl_msg_api_alloc_as_if_client_or_null (int nbytes) +{ + return vl_msg_api_alloc_internal (nbytes, 0, 1 /* may_return_null */ ); +} + +void +vl_msg_api_free (void *a) +{ + msgbuf_t *rv; + void *oldheap; + api_main_t *am = &api_main; + + rv = (msgbuf_t *) (((u8 *) a) - offsetof (msgbuf_t, data)); + + /* + * Here's the beauty of the scheme. Only one proc/thread has + * control of a given message buffer. To free a buffer, we just clear the + * queue field, and leave. No locks, no hits, no errors... + */ + if (rv->q) + { + rv->q = 0; + return; + } + + pthread_mutex_lock (&am->vlib_rp->mutex); + oldheap = svm_push_data_heap (am->vlib_rp); + clib_mem_free (rv); + svm_pop_heap (oldheap); + pthread_mutex_unlock (&am->vlib_rp->mutex); +} + +static void +vl_msg_api_free_nolock (void *a) +{ + msgbuf_t *rv; + void *oldheap; + api_main_t *am = &api_main; + + rv = (msgbuf_t *) (((u8 *) a) - offsetof (msgbuf_t, data)); + /* + * Here's the beauty of the scheme. Only one proc/thread has + * control of a given message buffer. To free a buffer, we just clear the + * queue field, and leave. No locks, no hits, no errors... + */ + if (rv->q) + { + rv->q = 0; + return; + } + + oldheap = svm_push_data_heap (am->vlib_rp); + clib_mem_free (rv); + svm_pop_heap (oldheap); +} + +void +vl_set_memory_root_path (char *name) +{ + api_main_t *am = &api_main; + + am->root_path = name; +} + +void +vl_set_memory_uid (int uid) +{ + api_main_t *am = &api_main; + + am->api_uid = uid; +} + +void +vl_set_memory_gid (int gid) +{ + api_main_t *am = &api_main; + + am->api_gid = gid; +} + +void +vl_set_global_memory_baseva (u64 baseva) +{ + api_main_t *am = &api_main; + + am->global_baseva = baseva; +} + +void +vl_set_global_memory_size (u64 size) +{ + api_main_t *am = &api_main; + + am->global_size = size; +} + +void +vl_set_api_memory_size (u64 size) +{ + api_main_t *am = &api_main; + + am->api_size = size; +} + +void +vl_set_global_pvt_heap_size (u64 size) +{ + api_main_t *am = &api_main; + + am->global_pvt_heap_size = size; +} + +void +vl_set_api_pvt_heap_size (u64 size) +{ + api_main_t *am = &api_main; + + am->api_pvt_heap_size = size; +} + +int +vl_map_shmem (char *region_name, int is_vlib) +{ + svm_map_region_args_t _a, *a = &_a; + svm_region_t *vlib_rp, *root_rp; + void *oldheap; + vl_shmem_hdr_t *shmem_hdr = 0; + api_main_t *am = &api_main; + int i; + struct timespec ts, tsrem; + + if (is_vlib == 0) + svm_region_init_chroot (am->root_path); + + memset (a, 0, sizeof (*a)); + + a->name = region_name; + a->size = am->api_size ? am->api_size : (16 << 20); + a->flags = SVM_FLAGS_MHEAP; + a->uid = am->api_uid; + a->gid = am->api_gid; + a->pvt_heap_size = am->api_pvt_heap_size; + + vlib_rp = svm_region_find_or_create (a); + + if (vlib_rp == 0) + return (-2); + + pthread_mutex_lock (&vlib_rp->mutex); + /* Has someone else set up the shared-memory variable table? */ + if (vlib_rp->user_ctx) + { + am->shmem_hdr = (void *) vlib_rp->user_ctx; + am->our_pid = getpid (); + if (is_vlib) + { + unix_shared_memory_queue_t *q; + uword old_msg; + /* + * application restart. Reset cached pids, API message + * rings, list of clients; otherwise, various things + * fail. (e.g. queue non-empty notification) + */ + + /* ghosts keep the region from disappearing properly */ + svm_client_scan_this_region_nolock (vlib_rp); + am->shmem_hdr->application_restarts++; + q = am->shmem_hdr->vl_input_queue; + am->shmem_hdr->vl_pid = getpid (); + q->consumer_pid = am->shmem_hdr->vl_pid; + /* Drain the input queue, freeing msgs */ + for (i = 0; i < 10; i++) + { + if (pthread_mutex_trylock (&q->mutex) == 0) + { + pthread_mutex_unlock (&q->mutex); + goto mutex_ok; + } + ts.tv_sec = 0; + ts.tv_nsec = 10000 * 1000; /* 10 ms */ + while (nanosleep (&ts, &tsrem) < 0) + ts = tsrem; + } + /* Mutex buggered, "fix" it */ + memset (&q->mutex, 0, sizeof (q->mutex)); + clib_warning ("forcibly release main input queue mutex"); + + mutex_ok: + am->vlib_rp = vlib_rp; + while (unix_shared_memory_queue_sub (q, + (u8 *) & old_msg, + 1 /* nowait */ ) + != -2 /* queue underflow */ ) + { + vl_msg_api_free_nolock ((void *) old_msg); + am->shmem_hdr->restart_reclaims++; + } + pthread_mutex_unlock (&vlib_rp->mutex); + root_rp = svm_get_root_rp (); + ASSERT (root_rp); + /* Clean up the root region client list */ + pthread_mutex_lock (&root_rp->mutex); + svm_client_scan_this_region_nolock (root_rp); + pthread_mutex_unlock (&root_rp->mutex); + } + else + { + pthread_mutex_unlock (&vlib_rp->mutex); + } + am->vlib_rp = vlib_rp; + vec_add1 (am->mapped_shmem_regions, vlib_rp); + return 0; + } + /* Clients simply have to wait... */ + if (!is_vlib) + { + pthread_mutex_unlock (&vlib_rp->mutex); + + /* Wait up to 100 seconds... */ + for (i = 0; i < 10000; i++) + { + ts.tv_sec = 0; + ts.tv_nsec = 10000 * 1000; /* 10 ms */ + while (nanosleep (&ts, &tsrem) < 0) + ts = tsrem; + if (vlib_rp->user_ctx) + goto ready; + } + /* Clean up and leave... */ + svm_region_unmap (vlib_rp); + clib_warning ("region init fail"); + return (-2); + + ready: + am->shmem_hdr = (void *) vlib_rp->user_ctx; + am->our_pid = getpid (); + am->vlib_rp = vlib_rp; + vec_add1 (am->mapped_shmem_regions, vlib_rp); + return 0; + } + + /* Nope, it's our problem... */ + + oldheap = svm_push_data_heap (vlib_rp); + + vec_validate (shmem_hdr, 0); + shmem_hdr->version = VL_SHM_VERSION; + + /* vlib main input queue */ + shmem_hdr->vl_input_queue = + unix_shared_memory_queue_init (1024, sizeof (uword), getpid (), + am->vlib_signal); + + /* Set up the msg ring allocator */ +#define _(sz,n) \ + do { \ + ring_alloc_t _rp; \ + _rp.rp = unix_shared_memory_queue_init ((n), (sz), 0, 0); \ + _rp.size = (sz); \ + _rp.nitems = n; \ + _rp.hits = 0; \ + _rp.misses = 0; \ + vec_add1(shmem_hdr->vl_rings, _rp); \ + } while (0); + + foreach_vl_aring_size; +#undef _ + +#define _(sz,n) \ + do { \ + ring_alloc_t _rp; \ + _rp.rp = unix_shared_memory_queue_init ((n), (sz), 0, 0); \ + _rp.size = (sz); \ + _rp.nitems = n; \ + _rp.hits = 0; \ + _rp.misses = 0; \ + vec_add1(shmem_hdr->client_rings, _rp); \ + } while (0); + + foreach_clnt_aring_size; +#undef _ + + am->shmem_hdr = shmem_hdr; + am->vlib_rp = vlib_rp; + am->our_pid = getpid (); + if (is_vlib) + am->shmem_hdr->vl_pid = am->our_pid; + + svm_pop_heap (oldheap); + + /* + * After absolutely everything that a client might see is set up, + * declare the shmem region valid + */ + vlib_rp->user_ctx = shmem_hdr; + + pthread_mutex_unlock (&vlib_rp->mutex); + vec_add1 (am->mapped_shmem_regions, vlib_rp); + return 0; +} + +void +vl_register_mapped_shmem_region (svm_region_t * rp) +{ + api_main_t *am = &api_main; + + vec_add1 (am->mapped_shmem_regions, rp); +} + +void +vl_unmap_shmem (void) +{ + svm_region_t *rp; + int i; + api_main_t *am = &api_main; + + if (!svm_get_root_rp ()) + return; + + for (i = 0; i < vec_len (am->mapped_shmem_regions); i++) + { + rp = am->mapped_shmem_regions[i]; + svm_region_unmap (rp); + } + + vec_free (am->mapped_shmem_regions); + am->shmem_hdr = 0; + + svm_region_exit (); + /* $$$ more careful cleanup, valgrind run... */ + vec_free (am->msg_handlers); + vec_free (am->msg_endian_handlers); + vec_free (am->msg_print_handlers); +} + +void +vl_msg_api_send_shmem (unix_shared_memory_queue_t * q, u8 * elem) +{ + api_main_t *am = &api_main; + uword *trace = (uword *) elem; + + if (am->tx_trace && am->tx_trace->enabled) + vl_msg_api_trace (am, am->tx_trace, (void *) trace[0]); + + (void) unix_shared_memory_queue_add (q, elem, 0 /* nowait */ ); +} + +void +vl_msg_api_send_shmem_nolock (unix_shared_memory_queue_t * q, u8 * elem) +{ + api_main_t *am = &api_main; + uword *trace = (uword *) elem; + + if (am->tx_trace && am->tx_trace->enabled) + vl_msg_api_trace (am, am->tx_trace, (void *) trace[0]); + + (void) unix_shared_memory_queue_add_nolock (q, elem); +} + +static void +vl_api_memclnt_create_reply_t_handler (vl_api_memclnt_create_reply_t * mp) +{ + serialize_main_t _sm, *sm = &_sm; + api_main_t *am = &api_main; + u8 *tblv; + u32 nmsgs; + int i; + u8 *name_and_crc; + u32 msg_index; + + am->my_client_index = mp->index; + am->my_registration = (vl_api_registration_t *) (uword) mp->handle; + + /* Clean out any previous hash table (unlikely) */ + if (am->msg_index_by_name_and_crc) + { + int i; + u8 **keys = 0; + hash_pair_t *hp; + /* *INDENT-OFF* */ + hash_foreach_pair (hp, am->msg_index_by_name_and_crc, + ({ + vec_add1 (keys, (u8 *) hp->key); + })); + /* *INDENT-ON* */ + for (i = 0; i < vec_len (keys); i++) + vec_free (keys[i]); + vec_free (keys); + } + + am->msg_index_by_name_and_crc = hash_create_string (0, sizeof (uword)); + + /* Recreate the vnet-side API message handler table */ + tblv = (u8 *) mp->message_table; + serialize_open_vector (sm, tblv); + unserialize_integer (sm, &nmsgs, sizeof (u32)); + + for (i = 0; i < nmsgs; i++) + { + msg_index = unserialize_likely_small_unsigned_integer (sm); + unserialize_cstring (sm, (char **) &name_and_crc); + hash_set_mem (am->msg_index_by_name_and_crc, name_and_crc, msg_index); + } +} + +u32 +vl_api_get_msg_index (u8 * name_and_crc) +{ + api_main_t *am = &api_main; + uword *p; + + if (am->msg_index_by_name_and_crc) + { + p = hash_get_mem (am->msg_index_by_name_and_crc, name_and_crc); + if (p) + return p[0]; + } + return ~0; +} + +int +vl_client_connect (char *name, int ctx_quota, int input_queue_size) +{ + svm_region_t *svm; + vl_api_memclnt_create_t *mp; + vl_api_memclnt_create_reply_t *rp; + unix_shared_memory_queue_t *vl_input_queue; + vl_shmem_hdr_t *shmem_hdr; + int rv = 0; + void *oldheap; + api_main_t *am = &api_main; + + if (am->my_registration) + { + clib_warning ("client %s already connected...", name); + return -1; + } + + if (am->vlib_rp == 0) + { + clib_warning ("am->vlib_rp NULL"); + return -1; + } + + svm = am->vlib_rp; + shmem_hdr = am->shmem_hdr; + + if (shmem_hdr == 0 || shmem_hdr->vl_input_queue == 0) + { + clib_warning ("shmem_hdr / input queue NULL"); + return -1; + } + + pthread_mutex_lock (&svm->mutex); + oldheap = svm_push_data_heap (svm); + vl_input_queue = + unix_shared_memory_queue_init (input_queue_size, sizeof (uword), + getpid (), 0); + pthread_mutex_unlock (&svm->mutex); + svm_pop_heap (oldheap); + + am->my_client_index = ~0; + am->my_registration = 0; + am->vl_input_queue = vl_input_queue; + + mp = vl_msg_api_alloc (sizeof (vl_api_memclnt_create_t)); + memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = ntohs (VL_API_MEMCLNT_CREATE); + mp->ctx_quota = ctx_quota; + mp->input_queue = (uword) vl_input_queue; + strncpy ((char *) mp->name, name, sizeof (mp->name) - 1); + + vl_msg_api_send_shmem (shmem_hdr->vl_input_queue, (u8 *) & mp); + + while (1) + { + int qstatus; + struct timespec ts, tsrem; + int i; + + /* Wait up to 10 seconds */ + for (i = 0; i < 1000; i++) + { + qstatus = unix_shared_memory_queue_sub (vl_input_queue, (u8 *) & rp, + 1 /* nowait */ ); + if (qstatus == 0) + goto read_one_msg; + ts.tv_sec = 0; + ts.tv_nsec = 10000 * 1000; /* 10 ms */ + while (nanosleep (&ts, &tsrem) < 0) + ts = tsrem; + } + /* Timeout... */ + clib_warning ("memclnt_create_reply timeout"); + return -1; + + read_one_msg: + if (ntohs (rp->_vl_msg_id) != VL_API_MEMCLNT_CREATE_REPLY) + { + clib_warning ("unexpected reply: id %d", ntohs (rp->_vl_msg_id)); + continue; + } + rv = clib_net_to_host_u32 (rp->response); + + vl_msg_api_handler ((void *) rp); + break; + } + return (rv); +} + +static void +vl_api_memclnt_delete_reply_t_handler (vl_api_memclnt_delete_reply_t * mp) +{ + void *oldheap; + api_main_t *am = &api_main; + + pthread_mutex_lock (&am->vlib_rp->mutex); + oldheap = svm_push_data_heap (am->vlib_rp); + unix_shared_memory_queue_free (am->vl_input_queue); + pthread_mutex_unlock (&am->vlib_rp->mutex); + svm_pop_heap (oldheap); + + am->my_client_index = ~0; + am->my_registration = 0; + am->vl_input_queue = 0; +} + +void +vl_client_disconnect (void) +{ + vl_api_memclnt_delete_t *mp; + vl_api_memclnt_delete_reply_t *rp; + unix_shared_memory_queue_t *vl_input_queue; + vl_shmem_hdr_t *shmem_hdr; + time_t begin; + api_main_t *am = &api_main; + + ASSERT (am->vlib_rp); + shmem_hdr = am->shmem_hdr; + ASSERT (shmem_hdr && shmem_hdr->vl_input_queue); + + vl_input_queue = am->vl_input_queue; + + mp = vl_msg_api_alloc (sizeof (vl_api_memclnt_delete_t)); + memset (mp, 0, sizeof (*mp)); + mp->_vl_msg_id = ntohs (VL_API_MEMCLNT_DELETE); + mp->index = am->my_client_index; + mp->handle = (uword) am->my_registration; + + vl_msg_api_send_shmem (shmem_hdr->vl_input_queue, (u8 *) & mp); + + /* + * Have to be careful here, in case the client is disconnecting + * because e.g. the vlib process died, or is unresponsive. + */ + + begin = time (0); + while (1) + { + time_t now; + + now = time (0); + + if (now >= (begin + 2)) + { + clib_warning ("peer unresponsive, give up"); + am->my_client_index = ~0; + am->my_registration = 0; + am->shmem_hdr = 0; + break; + } + if (unix_shared_memory_queue_sub (vl_input_queue, (u8 *) & rp, 1) < 0) + continue; + + /* drain the queue */ + if (ntohs (rp->_vl_msg_id) != VL_API_MEMCLNT_DELETE_REPLY) + { + vl_msg_api_handler ((void *) rp); + continue; + } + vl_msg_api_handler ((void *) rp); + break; + } +} + +static inline vl_api_registration_t * +vl_api_client_index_to_registration_internal (u32 handle) +{ + vl_api_registration_t **regpp; + vl_api_registration_t *regp; + api_main_t *am = &api_main; + u32 index; + + index = vl_msg_api_handle_get_index (handle); + if ((am->shmem_hdr->application_restarts & VL_API_EPOCH_MASK) + != vl_msg_api_handle_get_epoch (handle)) + { + vl_msg_api_increment_missing_client_counter (); + return 0; + } + + regpp = am->vl_clients + index; + + if (pool_is_free (am->vl_clients, regpp)) + { + vl_msg_api_increment_missing_client_counter (); + return 0; + } + regp = *regpp; + return (regp); +} + +vl_api_registration_t * +vl_api_client_index_to_registration (u32 index) +{ + return (vl_api_client_index_to_registration_internal (index)); +} + +unix_shared_memory_queue_t * +vl_api_client_index_to_input_queue (u32 index) +{ + vl_api_registration_t *regp; + api_main_t *am = &api_main; + + /* Special case: vlib trying to send itself a message */ + if (index == (u32) ~ 0) + return (am->shmem_hdr->vl_input_queue); + + regp = vl_api_client_index_to_registration_internal (index); + if (!regp) + return 0; + return (regp->vl_input_queue); +} + +#define foreach_api_client_msg \ +_(MEMCLNT_CREATE_REPLY, memclnt_create_reply) \ +_(MEMCLNT_DELETE_REPLY, memclnt_delete_reply) + +int +vl_client_api_map (char *region_name) +{ + int rv; + + if ((rv = vl_map_shmem (region_name, 0 /* is_vlib */ )) < 0) + { + return rv; + } + +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, 0 /* name */, \ + vl_api_##n##_t_handler, \ + 0/* cleanup */, 0/* endian */, 0/* print */, \ + sizeof(vl_api_##n##_t), 1); + foreach_api_client_msg; +#undef _ + return 0; +} + +void +vl_client_api_unmap (void) +{ + vl_unmap_shmem (); +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/memory_vlib.c b/src/vlibmemory/memory_vlib.c new file mode 100644 index 00000000000..1d40bcb791f --- /dev/null +++ b/src/vlibmemory/memory_vlib.c @@ -0,0 +1,1346 @@ +/* + *------------------------------------------------------------------ + * memory_vlib.c + * + * Copyright (c) 2009 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <unistd.h> +#include <sys/types.h> +#include <signal.h> +#include <pthread.h> +#include <vppinfra/vec.h> +#include <vppinfra/hash.h> +#include <vppinfra/pool.h> +#include <vppinfra/format.h> +#include <vppinfra/byte_order.h> +#include <vppinfra/elog.h> +#include <stdarg.h> +#include <vlib/vlib.h> +#include <vlib/unix/unix.h> +#include <vlibapi/api.h> +#include <vlibmemory/api.h> + +#define TRACE_VLIB_MEMORY_QUEUE 0 + +#include <vlibmemory/vl_memory_msg_enum.h> /* enumerate all vlib messages */ + +#define vl_typedefs /* define message structures */ +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_typedefs + +/* instantiate all the print functions we know about */ +#define vl_print(handle, ...) vlib_cli_output (handle, __VA_ARGS__) +#define vl_printfun +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_printfun + +static inline void * +vl_api_memclnt_create_t_print (vl_api_memclnt_create_t * a, void *handle) +{ + vl_print (handle, "vl_api_memclnt_create_t:\n"); + vl_print (handle, "name: %s\n", a->name); + vl_print (handle, "input_queue: 0x%wx\n", a->input_queue); + vl_print (handle, "context: %u\n", (unsigned) a->context); + vl_print (handle, "ctx_quota: %ld\n", (long) a->ctx_quota); + return handle; +} + +static inline void * +vl_api_memclnt_delete_t_print (vl_api_memclnt_delete_t * a, void *handle) +{ + vl_print (handle, "vl_api_memclnt_delete_t:\n"); + vl_print (handle, "index: %u\n", (unsigned) a->index); + vl_print (handle, "handle: 0x%wx\n", a->handle); + return handle; +} + +/* instantiate all the endian swap functions we know about */ +#define vl_endianfun +#include <vlibmemory/vl_memory_api_h.h> +#undef vl_endianfun + +void vl_socket_api_send (vl_api_registration_t * rp, u8 * elem) + __attribute__ ((weak)); + +void +vl_socket_api_send (vl_api_registration_t * rp, u8 * elem) +{ + static int count; + + if (count++ < 5) + clib_warning ("need to link against -lvlibsocket, msg not sent!"); +} + +void +vl_msg_api_send (vl_api_registration_t * rp, u8 * elem) +{ + if (PREDICT_FALSE (rp->registration_type > REGISTRATION_TYPE_SHMEM)) + { + vl_socket_api_send (rp, elem); + } + else + { + vl_msg_api_send_shmem (rp->vl_input_queue, elem); + } +} + +u8 * +vl_api_serialize_message_table (api_main_t * am, u8 * vector) +{ + serialize_main_t _sm, *sm = &_sm; + hash_pair_t *hp; + u32 nmsg = hash_elts (am->msg_index_by_name_and_crc); + + serialize_open_vector (sm, vector); + + /* serialize the count */ + serialize_integer (sm, nmsg, sizeof (u32)); + + hash_foreach_pair (hp, am->msg_index_by_name_and_crc, ( + { + serialize_likely_small_unsigned_integer + (sm, hp->value[0]); + serialize_cstring + (sm, + (char *) hp->key); + })); + + return serialize_close_vector (sm); +} + +/* + * vl_api_memclnt_create_t_handler + */ +void +vl_api_memclnt_create_t_handler (vl_api_memclnt_create_t * mp) +{ + vl_api_registration_t **regpp; + vl_api_registration_t *regp; + vl_api_memclnt_create_reply_t *rp; + svm_region_t *svm; + unix_shared_memory_queue_t *q; + int rv = 0; + void *oldheap; + api_main_t *am = &api_main; + u8 *serialized_message_table = 0; + + /* + * This is tortured. Maintain a vlib-address-space private + * pool of client registrations. We use the shared-memory virtual + * address of client structure as a handle, to allow direct + * manipulation of context quota vbls from the client library. + * + * This scheme causes trouble w/ API message trace replay, since + * some random VA from clib_mem_alloc() certainly won't + * occur in the Linux sim. The (very) few places + * that care need to use the pool index. + * + * Putting the registration object(s) into a pool in shared memory and + * using the pool index as a handle seems like a great idea. + * Unfortunately, each and every reference to that pool would need + * to be protected by a mutex: + * + * Client VLIB + * ------ ---- + * convert pool index to + * pointer. + * <deschedule> + * expand pool + * <deschedule> + * kaboom! + */ + + pool_get (am->vl_clients, regpp); + + svm = am->vlib_rp; + + if (am->serialized_message_table_in_shmem == 0) + serialized_message_table = vl_api_serialize_message_table (am, 0); + + pthread_mutex_lock (&svm->mutex); + oldheap = svm_push_data_heap (svm); + *regpp = clib_mem_alloc (sizeof (vl_api_registration_t)); + + regp = *regpp; + memset (regp, 0, sizeof (*regp)); + regp->registration_type = REGISTRATION_TYPE_SHMEM; + regp->vl_api_registration_pool_index = regpp - am->vl_clients; + + q = regp->vl_input_queue = (unix_shared_memory_queue_t *) (uword) + mp->input_queue; + + regp->name = format (0, "%s", mp->name); + vec_add1 (regp->name, 0); + if (serialized_message_table) + am->serialized_message_table_in_shmem = + vec_dup (serialized_message_table); + + pthread_mutex_unlock (&svm->mutex); + svm_pop_heap (oldheap); + + vec_free (serialized_message_table); + + rp = vl_msg_api_alloc (sizeof (*rp)); + rp->_vl_msg_id = ntohs (VL_API_MEMCLNT_CREATE_REPLY); + rp->handle = (uword) regp; + rp->index = vl_msg_api_handle_from_index_and_epoch + (regp->vl_api_registration_pool_index, + am->shmem_hdr->application_restarts); + rp->context = mp->context; + rp->response = ntohl (rv); + rp->message_table = (u64) am->serialized_message_table_in_shmem; + + vl_msg_api_send_shmem (q, (u8 *) & rp); +} + +/* Application callback to clean up leftover registrations from this client */ +int vl_api_memclnt_delete_callback (u32 client_index) __attribute__ ((weak)); + +int +vl_api_memclnt_delete_callback (u32 client_index) +{ + return 0; +} + +/* + * vl_api_memclnt_delete_t_handler + */ +void +vl_api_memclnt_delete_t_handler (vl_api_memclnt_delete_t * mp) +{ + vl_api_registration_t **regpp; + vl_api_registration_t *regp; + vl_api_memclnt_delete_reply_t *rp; + svm_region_t *svm; + void *oldheap; + api_main_t *am = &api_main; + u32 handle, client_index, epoch; + + handle = mp->index; + + if (vl_api_memclnt_delete_callback (handle)) + return; + + epoch = vl_msg_api_handle_get_epoch (handle); + client_index = vl_msg_api_handle_get_index (handle); + + if (epoch != (am->shmem_hdr->application_restarts & VL_API_EPOCH_MASK)) + { + clib_warning + ("Stale clnt delete index %d old epoch %d cur epoch %d", + client_index, epoch, + (am->shmem_hdr->application_restarts & VL_API_EPOCH_MASK)); + return; + } + + regpp = am->vl_clients + client_index; + + if (!pool_is_free (am->vl_clients, regpp)) + { + regp = *regpp; + svm = am->vlib_rp; + + /* $$$ check the input queue for e.g. punted sf's */ + + rp = vl_msg_api_alloc (sizeof (*rp)); + rp->_vl_msg_id = ntohs (VL_API_MEMCLNT_DELETE_REPLY); + rp->handle = mp->handle; + rp->response = 1; + + vl_msg_api_send_shmem (regp->vl_input_queue, (u8 *) & rp); + + if (client_index != regp->vl_api_registration_pool_index) + { + clib_warning ("mismatch client_index %d pool_index %d", + client_index, regp->vl_api_registration_pool_index); + vl_msg_api_free (rp); + return; + } + + /* No dangling references, please */ + *regpp = 0; + + pool_put_index (am->vl_clients, regp->vl_api_registration_pool_index); + + pthread_mutex_lock (&svm->mutex); + oldheap = svm_push_data_heap (svm); + /* Poison the old registration */ + memset (regp, 0xF1, sizeof (*regp)); + clib_mem_free (regp); + pthread_mutex_unlock (&svm->mutex); + svm_pop_heap (oldheap); + } + else + { + clib_warning ("unknown client ID %d", mp->index); + } +} + +void +vl_api_get_first_msg_id_t_handler (vl_api_get_first_msg_id_t * mp) +{ + vl_api_get_first_msg_id_reply_t *rmp; + unix_shared_memory_queue_t *q; + uword *p; + api_main_t *am = &api_main; + vl_api_msg_range_t *rp; + u8 name[64]; + u16 first_msg_id = ~0; + int rv = -7; /* VNET_API_ERROR_INVALID_VALUE */ + + q = vl_api_client_index_to_input_queue (mp->client_index); + if (!q) + return; + + if (am->msg_range_by_name == 0) + goto out; + + strncpy ((char *) name, (char *) mp->name, ARRAY_LEN (name) - 1); + + p = hash_get_mem (am->msg_range_by_name, name); + if (p == 0) + goto out; + + rp = vec_elt_at_index (am->msg_ranges, p[0]); + + first_msg_id = rp->first_msg_id; + rv = 0; + +out: + + rmp = vl_msg_api_alloc (sizeof (*rmp)); + rmp->_vl_msg_id = ntohs (VL_API_GET_FIRST_MSG_ID_REPLY); + rmp->context = mp->context; + rmp->retval = ntohl (rv); + rmp->first_msg_id = ntohs (first_msg_id); + vl_msg_api_send_shmem (q, (u8 *) & rmp); +} + +#define foreach_vlib_api_msg \ +_(MEMCLNT_CREATE, memclnt_create) \ +_(MEMCLNT_DELETE, memclnt_delete) \ +_(GET_FIRST_MSG_ID, get_first_msg_id) + +/* + * vl_api_init + */ +static int +memory_api_init (char *region_name) +{ + int rv; + vl_msg_api_msg_config_t cfg; + vl_msg_api_msg_config_t *c = &cfg; + + if ((rv = vl_map_shmem (region_name, 1 /* is_vlib */ )) < 0) + return rv; + +#define _(N,n) do { \ + c->id = VL_API_##N; \ + c->name = #n; \ + c->handler = vl_api_##n##_t_handler; \ + c->cleanup = vl_noop_handler; \ + c->endian = vl_api_##n##_t_endian; \ + c->print = vl_api_##n##_t_print; \ + c->size = sizeof(vl_api_##n##_t); \ + c->traced = 1; /* trace, so these msgs print */ \ + c->replay = 0; /* don't replay client create/delete msgs */ \ + vl_msg_api_config(c);} while (0); + + foreach_vlib_api_msg; +#undef _ + + return 0; +} + +#define foreach_histogram_bucket \ +_(400) \ +_(200) \ +_(100) \ +_(10) + +typedef enum +{ +#define _(n) SLEEP_##n##_US, + foreach_histogram_bucket +#undef _ + SLEEP_N_BUCKETS, +} histogram_index_t; + +static u64 vector_rate_histogram[SLEEP_N_BUCKETS]; + +static void memclnt_queue_callback (vlib_main_t * vm); + +static uword +memclnt_process (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * f) +{ + uword mp; + vl_shmem_hdr_t *shm; + unix_shared_memory_queue_t *q; + clib_error_t *e; + int rv; + api_main_t *am = &api_main; + f64 dead_client_scan_time; + f64 sleep_time, start_time; + f64 vector_rate; + + vlib_set_queue_signal_callback (vm, memclnt_queue_callback); + + if ((rv = memory_api_init (am->region_name)) < 0) + { + clib_warning ("memory_api_init returned %d, wait for godot...", rv); + vlib_process_suspend (vm, 1e70); + } + + shm = am->shmem_hdr; + ASSERT (shm); + q = shm->vl_input_queue; + ASSERT (q); + + e = vlib_call_init_exit_functions + (vm, vm->api_init_function_registrations, 1 /* call_once */ ); + if (e) + clib_error_report (e); + + sleep_time = 20.0; + dead_client_scan_time = vlib_time_now (vm) + 20.0; + + /* $$$ pay attention to frame size, control CPU usage */ + while (1) + { + uword event_type __attribute__ ((unused)); + i8 *headp; + int need_broadcast; + + /* + * There's a reason for checking the queue before + * sleeping. If the vlib application crashes, it's entirely + * possible for a client to enqueue a connect request + * during the process restart interval. + * + * Unless some force of physics causes the new incarnation + * of the application to process the request, the client will + * sit and wait for Godot... + */ + vector_rate = vlib_last_vector_length_per_node (vm); + start_time = vlib_time_now (vm); + while (1) + { + pthread_mutex_lock (&q->mutex); + if (q->cursize == 0) + { + vm->api_queue_nonempty = 0; + pthread_mutex_unlock (&q->mutex); + + if (TRACE_VLIB_MEMORY_QUEUE) + { + /* *INDENT-OFF* */ + ELOG_TYPE_DECLARE (e) = + { + .format = "q-underflow: len %d", + .format_args = "i4", + }; + /* *INDENT-ON* */ + struct + { + u32 len; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->len = 0; + } + sleep_time = 20.0; + break; + } + + headp = (i8 *) (q->data + sizeof (uword) * q->head); + clib_memcpy (&mp, headp, sizeof (uword)); + + q->head++; + need_broadcast = (q->cursize == q->maxsize / 2); + q->cursize--; + + if (PREDICT_FALSE (q->head == q->maxsize)) + q->head = 0; + pthread_mutex_unlock (&q->mutex); + if (need_broadcast) + (void) pthread_cond_broadcast (&q->condvar); + + vl_msg_api_handler_with_vm_node (am, (void *) mp, vm, node); + + /* Allow no more than 10us without a pause */ + if (vlib_time_now (vm) > start_time + 10e-6) + { + int index = SLEEP_400_US; + if (vector_rate > 40.0) + sleep_time = 400e-6; + else if (vector_rate > 20.0) + { + index = SLEEP_200_US; + sleep_time = 200e-6; + } + else if (vector_rate >= 1.0) + { + index = SLEEP_100_US; + sleep_time = 100e-6; + } + else + { + index = SLEEP_10_US; + sleep_time = 10e-6; + } + vector_rate_histogram[index] += 1; + break; + } + } + + event_type = vlib_process_wait_for_event_or_clock (vm, sleep_time); + vm->queue_signal_pending = 0; + vlib_process_get_events (vm, 0 /* event_data */ ); + + if (vlib_time_now (vm) > dead_client_scan_time) + { + vl_api_registration_t **regpp; + vl_api_registration_t *regp; + unix_shared_memory_queue_t *q; + static u32 *dead_indices; + static u32 *confused_indices; + + vec_reset_length (dead_indices); + vec_reset_length (confused_indices); + + /* *INDENT-OFF* */ + pool_foreach (regpp, am->vl_clients, + ({ + regp = *regpp; + if (regp) + { + q = regp->vl_input_queue; + if (kill (q->consumer_pid, 0) < 0) + { + vec_add1(dead_indices, regpp - am->vl_clients); + } + } + else + { + clib_warning ("NULL client registration index %d", + regpp - am->vl_clients); + vec_add1 (confused_indices, regpp - am->vl_clients); + } + })); + /* *INDENT-ON* */ + /* This should "never happen," but if it does, fix it... */ + if (PREDICT_FALSE (vec_len (confused_indices) > 0)) + { + int i; + for (i = 0; i < vec_len (confused_indices); i++) + { + pool_put_index (am->vl_clients, confused_indices[i]); + } + } + + if (PREDICT_FALSE (vec_len (dead_indices) > 0)) + { + int i; + svm_region_t *svm; + void *oldheap; + + /* Allow the application to clean up its registrations */ + for (i = 0; i < vec_len (dead_indices); i++) + { + regpp = pool_elt_at_index (am->vl_clients, dead_indices[i]); + if (regpp) + { + u32 handle; + + handle = vl_msg_api_handle_from_index_and_epoch + (dead_indices[i], shm->application_restarts); + (void) vl_api_memclnt_delete_callback (handle); + } + } + + svm = am->vlib_rp; + pthread_mutex_lock (&svm->mutex); + oldheap = svm_push_data_heap (svm); + + for (i = 0; i < vec_len (dead_indices); i++) + { + regpp = pool_elt_at_index (am->vl_clients, dead_indices[i]); + if (regpp) + { + /* Poison the old registration */ + memset (*regpp, 0xF3, sizeof (**regpp)); + clib_mem_free (*regpp); + /* no dangling references, please */ + *regpp = 0; + } + else + { + svm_pop_heap (oldheap); + clib_warning ("Duplicate free, client index %d", + regpp - am->vl_clients); + oldheap = svm_push_data_heap (svm); + } + } + + svm_client_scan_this_region_nolock (am->vlib_rp); + + pthread_mutex_unlock (&svm->mutex); + svm_pop_heap (oldheap); + for (i = 0; i < vec_len (dead_indices); i++) + pool_put_index (am->vl_clients, dead_indices[i]); + } + + dead_client_scan_time = vlib_time_now (vm) + 20.0; + } + + if (TRACE_VLIB_MEMORY_QUEUE) + { + /* *INDENT-OFF* */ + ELOG_TYPE_DECLARE (e) = { + .format = "q-awake: len %d", + .format_args = "i4", + }; + /* *INDENT-ON* */ + struct + { + u32 len; + } *ed; + ed = ELOG_DATA (&vm->elog_main, e); + ed->len = q->cursize; + } + } + + return 0; +} + +static clib_error_t * +vl_api_show_histogram_command (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cli_cmd) +{ + u64 total_counts = 0; + int i; + + for (i = 0; i < SLEEP_N_BUCKETS; i++) + { + total_counts += vector_rate_histogram[i]; + } + + if (total_counts == 0) + { + vlib_cli_output (vm, "No control-plane activity."); + return 0; + } + +#define _(n) \ + do { \ + f64 percent; \ + percent = ((f64) vector_rate_histogram[SLEEP_##n##_US]) \ + / (f64) total_counts; \ + percent *= 100.0; \ + vlib_cli_output (vm, "Sleep %3d us: %llu, %.2f%%",n, \ + vector_rate_histogram[SLEEP_##n##_US], \ + percent); \ + } while (0); + foreach_histogram_bucket; +#undef _ + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_histogram_command, static) = { + .path = "show api histogram", + .short_help = "show api histogram", + .function = vl_api_show_histogram_command, +}; +/* *INDENT-ON* */ + +static clib_error_t * +vl_api_clear_histogram_command (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cli_cmd) +{ + int i; + + for (i = 0; i < SLEEP_N_BUCKETS; i++) + vector_rate_histogram[i] = 0; + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_clear_api_histogram_command, static) = { + .path = "clear api histogram", + .short_help = "clear api histogram", + .function = vl_api_clear_histogram_command, +}; +/* *INDENT-ON* */ + + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (memclnt_node,static) = { + .function = memclnt_process, + .type = VLIB_NODE_TYPE_PROCESS, + .name = "api-rx-from-ring", + .state = VLIB_NODE_STATE_DISABLED, +}; +/* *INDENT-ON* */ + +static void +memclnt_queue_callback (vlib_main_t * vm) +{ + static volatile int *cursizep; + + if (PREDICT_FALSE (cursizep == 0)) + { + api_main_t *am = &api_main; + vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr; + unix_shared_memory_queue_t *q; + + if (shmem_hdr == 0) + return; + + q = shmem_hdr->vl_input_queue; + if (q == 0) + return; + cursizep = &q->cursize; + } + + if (*cursizep >= 1) + { + vm->queue_signal_pending = 1; + vm->api_queue_nonempty = 1; + vlib_process_signal_event (vm, memclnt_node.index, + /* event_type */ 0, /* event_data */ 0); + } +} + +void +vl_enable_disable_memory_api (vlib_main_t * vm, int enable) +{ + vlib_node_set_state (vm, memclnt_node.index, + (enable + ? VLIB_NODE_STATE_POLLING + : VLIB_NODE_STATE_DISABLED)); +} + +static uword +api_rx_from_node (vlib_main_t * vm, + vlib_node_runtime_t * node, vlib_frame_t * frame) +{ + uword n_packets = frame->n_vectors; + uword n_left_from; + u32 *from; + static u8 *long_msg; + + vec_validate (long_msg, 4095); + n_left_from = frame->n_vectors; + from = vlib_frame_args (frame); + + while (n_left_from > 0) + { + u32 bi0; + vlib_buffer_t *b0; + void *msg; + uword msg_len; + + bi0 = from[0]; + b0 = vlib_get_buffer (vm, bi0); + from += 1; + n_left_from -= 1; + + msg = b0->data + b0->current_data; + msg_len = b0->current_length; + if (b0->flags & VLIB_BUFFER_NEXT_PRESENT) + { + ASSERT (long_msg != 0); + _vec_len (long_msg) = 0; + vec_add (long_msg, msg, msg_len); + while (b0->flags & VLIB_BUFFER_NEXT_PRESENT) + { + b0 = vlib_get_buffer (vm, b0->next_buffer); + msg = b0->data + b0->current_data; + msg_len = b0->current_length; + vec_add (long_msg, msg, msg_len); + } + msg = long_msg; + } + vl_msg_api_handler_no_trace_no_free (msg); + } + + /* Free what we've been given. */ + vlib_buffer_free (vm, vlib_frame_args (frame), n_packets); + + return n_packets; +} + +/* *INDENT-OFF* */ +VLIB_REGISTER_NODE (api_rx_from_node_node,static) = { + .function = api_rx_from_node, + .type = VLIB_NODE_TYPE_INTERNAL, + .vector_size = 4, + .name = "api-rx-from-node", +}; +/* *INDENT-ON* */ + +static clib_error_t * +setup_memclnt_exit (vlib_main_t * vm) +{ + atexit (vl_unmap_shmem); + return 0; +} + +VLIB_INIT_FUNCTION (setup_memclnt_exit); + + +static clib_error_t * +vl_api_ring_command (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cli_cmd) +{ + int i; + ring_alloc_t *ap; + vl_shmem_hdr_t *shmem_hdr; + api_main_t *am = &api_main; + + shmem_hdr = am->shmem_hdr; + + if (shmem_hdr == 0) + { + vlib_cli_output (vm, "Shared memory segment not initialized...\n"); + return 0; + } + + vlib_cli_output (vm, "%8s %8s %8s %8s %8s\n", + "Owner", "Size", "Nitems", "Hits", "Misses"); + + ap = shmem_hdr->vl_rings; + + for (i = 0; i < vec_len (shmem_hdr->vl_rings); i++) + { + vlib_cli_output (vm, "%8s %8d %8d %8d %8d\n", + "vlib", ap->size, ap->nitems, ap->hits, ap->misses); + ap++; + } + + ap = shmem_hdr->client_rings; + + for (i = 0; i < vec_len (shmem_hdr->client_rings); i++) + { + vlib_cli_output (vm, "%8s %8d %8d %8d %8d\n", + "clnt", ap->size, ap->nitems, ap->hits, ap->misses); + ap++; + } + + vlib_cli_output (vm, "%d ring miss fallback allocations\n", + am->ring_misses); + + vlib_cli_output (vm, "%d application restarts, %d reclaimed msgs\n", + shmem_hdr->application_restarts, + shmem_hdr->restart_reclaims); + return 0; +} + +void dump_socket_clients (vlib_main_t * vm, api_main_t * am) + __attribute__ ((weak)); + +void +dump_socket_clients (vlib_main_t * vm, api_main_t * am) +{ +} + +static clib_error_t * +vl_api_client_command (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cli_cmd) +{ + vl_api_registration_t **regpp, *regp; + unix_shared_memory_queue_t *q; + char *health; + api_main_t *am = &api_main; + u32 *confused_indices = 0; + + if (!pool_elts (am->vl_clients)) + goto socket_clients; + vlib_cli_output (vm, "Shared memory clients"); + vlib_cli_output (vm, "%16s %8s %14s %18s %s", + "Name", "PID", "Queue Length", "Queue VA", "Health"); + + /* *INDENT-OFF* */ + pool_foreach (regpp, am->vl_clients, + ({ + regp = *regpp; + + if (regp) + { + q = regp->vl_input_queue; + if (kill (q->consumer_pid, 0) < 0) + { + health = "DEAD"; + } + else + { + health = "alive"; + } + vlib_cli_output (vm, "%16s %8d %14d 0x%016llx %s\n", + regp->name, q->consumer_pid, q->cursize, + q, health); + } + else + { + clib_warning ("NULL client registration index %d", + regpp - am->vl_clients); + vec_add1 (confused_indices, regpp - am->vl_clients); + } + })); + /* *INDENT-ON* */ + + /* This should "never happen," but if it does, fix it... */ + if (PREDICT_FALSE (vec_len (confused_indices) > 0)) + { + int i; + for (i = 0; i < vec_len (confused_indices); i++) + { + pool_put_index (am->vl_clients, confused_indices[i]); + } + } + vec_free (confused_indices); + + if (am->missing_clients) + vlib_cli_output (vm, "%u messages with missing clients", + am->missing_clients); +socket_clients: + dump_socket_clients (vm, am); + + return 0; +} + +static clib_error_t * +vl_api_status_command (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cli_cmd) +{ + api_main_t *am = &api_main; + + // check if rx_trace and tx_trace are not null pointers + + if (am->rx_trace == 0) + { + vlib_cli_output (vm, "RX Trace disabled\n"); + } + else + { + if (am->rx_trace->enabled == 0) + vlib_cli_output (vm, "RX Trace disabled\n"); + else + vlib_cli_output (vm, "RX Trace enabled\n"); + } + + if (am->tx_trace == 0) + { + vlib_cli_output (vm, "TX Trace disabled\n"); + } + else + { + if (am->tx_trace->enabled == 0) + vlib_cli_output (vm, "TX Trace disabled\n"); + else + vlib_cli_output (vm, "TX Trace enabled\n"); + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_command, static) = { + .path = "show api", + .short_help = "Show API information", +}; +/* *INDENT-ON* */ + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_ring_command, static) = { + .path = "show api ring-stats", + .short_help = "Message ring statistics", + .function = vl_api_ring_command, +}; +/* *INDENT-ON* */ + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_clients_command, static) = { + .path = "show api clients", + .short_help = "Client information", + .function = vl_api_client_command, +}; +/* *INDENT-ON* */ + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_status_command, static) = { + .path = "show api status", + .short_help = "Show API trace status", + .function = vl_api_status_command, +}; +/* *INDENT-ON* */ + +static clib_error_t * +vl_api_message_table_command (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cli_cmd) +{ + api_main_t *am = &api_main; + int i; + int verbose = 0; + + if (unformat (input, "verbose")) + verbose = 1; + + + if (verbose == 0) + vlib_cli_output (vm, "%-4s %s", "ID", "Name"); + else + vlib_cli_output (vm, "%-4s %-40s %6s %7s", "ID", "Name", "Bounce", + "MP-safe"); + + for (i = 1; i < vec_len (am->msg_names); i++) + { + if (verbose == 0) + { + vlib_cli_output (vm, "%-4d %s", i, + am->msg_names[i] ? am->msg_names[i] : + " [no handler]"); + } + else + { + vlib_cli_output (vm, "%-4d %-40s %6d %7d", i, + am->msg_names[i] ? am->msg_names[i] : + " [no handler]", am->message_bounce[i], + am->is_mp_safe[i]); + } + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_message_table_command, static) = { + .path = "show api message-table", + .short_help = "Message Table", + .function = vl_api_message_table_command, +}; +/* *INDENT-ON* */ + +static clib_error_t * +vl_api_trace_command (vlib_main_t * vm, + unformat_input_t * input, vlib_cli_command_t * cli_cmd) +{ + u32 nitems = 1024; + vl_api_trace_which_t which = VL_API_TRACE_RX; + api_main_t *am = &api_main; + + while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT) + { + if (unformat (input, "rx nitems %u", &nitems) || unformat (input, "rx")) + goto configure; + else if (unformat (input, "tx nitems %u", &nitems) + || unformat (input, "tx")) + { + which = VL_API_TRACE_RX; + goto configure; + } + else if (unformat (input, "on rx")) + { + vl_msg_api_trace_onoff (am, VL_API_TRACE_RX, 1); + } + else if (unformat (input, "on tx")) + { + vl_msg_api_trace_onoff (am, VL_API_TRACE_TX, 1); + } + else if (unformat (input, "on")) + { + vl_msg_api_trace_onoff (am, VL_API_TRACE_RX, 1); + } + else if (unformat (input, "off")) + { + vl_msg_api_trace_onoff (am, VL_API_TRACE_RX, 0); + vl_msg_api_trace_onoff (am, VL_API_TRACE_TX, 0); + } + else if (unformat (input, "free")) + { + vl_msg_api_trace_onoff (am, VL_API_TRACE_RX, 0); + vl_msg_api_trace_onoff (am, VL_API_TRACE_TX, 0); + vl_msg_api_trace_free (am, VL_API_TRACE_RX); + vl_msg_api_trace_free (am, VL_API_TRACE_TX); + } + else if (unformat (input, "debug on")) + { + am->msg_print_flag = 1; + } + else if (unformat (input, "debug off")) + { + am->msg_print_flag = 0; + } + else + return clib_error_return (0, "unknown input `%U'", + format_unformat_error, input); + } + return 0; + +configure: + if (vl_msg_api_trace_configure (am, which, nitems)) + { + vlib_cli_output (vm, "warning: trace configure error (%d, %d)", + which, nitems); + } + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (trace, static) = { + .path = "set api-trace", + .short_help = "API trace", + .function = vl_api_trace_command, +}; +/* *INDENT-ON* */ + +clib_error_t * +vlibmemory_init (vlib_main_t * vm) +{ + api_main_t *am = &api_main; + svm_map_region_args_t _a, *a = &_a; + + memset (a, 0, sizeof (*a)); + a->root_path = am->root_path; + a->name = SVM_GLOBAL_REGION_NAME; + a->baseva = (am->global_baseva != 0) ? + am->global_baseva : SVM_GLOBAL_REGION_BASEVA; + a->size = (am->global_size != 0) ? am->global_size : SVM_GLOBAL_REGION_SIZE; + a->flags = SVM_FLAGS_NODATA; + a->uid = am->api_uid; + a->gid = am->api_gid; + a->pvt_heap_size = + (am->global_pvt_heap_size != + 0) ? am->global_pvt_heap_size : SVM_PVT_MHEAP_SIZE; + + svm_region_init_args (a); + return 0; +} + +VLIB_INIT_FUNCTION (vlibmemory_init); + +void +vl_set_memory_region_name (char *name) +{ + api_main_t *am = &api_main; + + am->region_name = name; +} + +static int +range_compare (vl_api_msg_range_t * a0, vl_api_msg_range_t * a1) +{ + int len0, len1, clen; + + len0 = vec_len (a0->name); + len1 = vec_len (a1->name); + clen = len0 < len1 ? len0 : len1; + return (strncmp ((char *) a0->name, (char *) a1->name, clen)); +} + +static u8 * +format_api_msg_range (u8 * s, va_list * args) +{ + vl_api_msg_range_t *rp = va_arg (*args, vl_api_msg_range_t *); + + if (rp == 0) + s = format (s, "%-20s%9s%9s", "Name", "First-ID", "Last-ID"); + else + s = format (s, "%-20s%9d%9d", rp->name, rp->first_msg_id, + rp->last_msg_id); + + return s; +} + +static clib_error_t * +vl_api_show_plugin_command (vlib_main_t * vm, + unformat_input_t * input, + vlib_cli_command_t * cli_cmd) +{ + api_main_t *am = &api_main; + vl_api_msg_range_t *rp = 0; + int i; + + if (vec_len (am->msg_ranges) == 0) + { + vlib_cli_output (vm, "No plugin API message ranges configured..."); + return 0; + } + + rp = vec_dup (am->msg_ranges); + + vec_sort_with_function (rp, range_compare); + + vlib_cli_output (vm, "Plugin API message ID ranges...\n"); + vlib_cli_output (vm, "%U", format_api_msg_range, 0 /* header */ ); + + for (i = 0; i < vec_len (rp); i++) + vlib_cli_output (vm, "%U", format_api_msg_range, rp + i); + + return 0; +} + +/* *INDENT-OFF* */ +VLIB_CLI_COMMAND (cli_show_api_plugin_command, static) = { + .path = "show api plugin", + .short_help = "show api plugin", + .function = vl_api_show_plugin_command, +}; +/* *INDENT-ON* */ + +static void +vl_api_rpc_call_t_handler (vl_api_rpc_call_t * mp) +{ + vl_api_rpc_reply_t *rmp; + int (*fp) (void *); + i32 rv = 0; + vlib_main_t *vm = vlib_get_main (); + + if (mp->function == 0) + { + rv = -1; + clib_warning ("rpc NULL function pointer"); + } + + else + { + if (mp->need_barrier_sync) + vlib_worker_thread_barrier_sync (vm); + + fp = uword_to_pointer (mp->function, int (*)(void *)); + rv = fp (mp->data); + + if (mp->need_barrier_sync) + vlib_worker_thread_barrier_release (vm); + } + + if (mp->send_reply) + { + unix_shared_memory_queue_t *q = + vl_api_client_index_to_input_queue (mp->client_index); + if (q) + { + rmp = vl_msg_api_alloc_as_if_client (sizeof (*rmp)); + rmp->_vl_msg_id = ntohs (VL_API_RPC_REPLY); + rmp->context = mp->context; + rmp->retval = rv; + vl_msg_api_send_shmem (q, (u8 *) & rmp); + } + } + if (mp->multicast) + { + clib_warning ("multicast not yet implemented..."); + } +} + +static void +vl_api_rpc_reply_t_handler (vl_api_rpc_reply_t * mp) +{ + clib_warning ("unimplemented"); +} + +void +vl_api_rpc_call_main_thread (void *fp, u8 * data, u32 data_length) +{ + vl_api_rpc_call_t *mp; + api_main_t *am = &api_main; + vl_shmem_hdr_t *shmem_hdr = am->shmem_hdr; + unix_shared_memory_queue_t *q; + + /* Main thread: call the function directly */ + if (os_get_cpu_number () == 0) + { + vlib_main_t *vm = vlib_get_main (); + void (*call_fp) (void *); + + vlib_worker_thread_barrier_sync (vm); + + call_fp = fp; + call_fp (data); + + vlib_worker_thread_barrier_release (vm); + return; + } + + /* Any other thread, actually do an RPC call... */ + mp = vl_msg_api_alloc_as_if_client (sizeof (*mp) + data_length); + + memset (mp, 0, sizeof (*mp)); + clib_memcpy (mp->data, data, data_length); + mp->_vl_msg_id = ntohs (VL_API_RPC_CALL); + mp->function = pointer_to_uword (fp); + mp->need_barrier_sync = 1; + + /* + * Use the "normal" control-plane mechanism for the main thread. + * Well, almost. if the main input queue is full, we cannot + * block. Otherwise, we can expect a barrier sync timeout. + */ + q = shmem_hdr->vl_input_queue; + + while (pthread_mutex_trylock (&q->mutex)) + vlib_worker_thread_barrier_check (); + + while (PREDICT_FALSE (unix_shared_memory_queue_is_full (q))) + { + pthread_mutex_unlock (&q->mutex); + vlib_worker_thread_barrier_check (); + while (pthread_mutex_trylock (&q->mutex)) + vlib_worker_thread_barrier_check (); + } + + vl_msg_api_send_shmem_nolock (q, (u8 *) & mp); + + pthread_mutex_unlock (&q->mutex); +} + +#define foreach_rpc_api_msg \ +_(RPC_CALL,rpc_call) \ +_(RPC_REPLY,rpc_reply) + +static clib_error_t * +rpc_api_hookup (vlib_main_t * vm) +{ +#define _(N,n) \ + vl_msg_api_set_handlers(VL_API_##N, #n, \ + vl_api_##n##_t_handler, \ + vl_noop_handler, \ + vl_noop_handler, \ + vl_api_##n##_t_print, \ + sizeof(vl_api_##n##_t), 0 /* do not trace */); + foreach_rpc_api_msg; +#undef _ + return 0; +} + +VLIB_API_INIT_FUNCTION (rpc_api_hookup); + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/unix_shared_memory_queue.c b/src/vlibmemory/unix_shared_memory_queue.c new file mode 100644 index 00000000000..25d28910706 --- /dev/null +++ b/src/vlibmemory/unix_shared_memory_queue.c @@ -0,0 +1,324 @@ +/* + *------------------------------------------------------------------ + * unix_shared_memory_queue.c - unidirectional shared-memory queues + * + * Copyright (c) 2009 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <pthread.h> +#include <vppinfra/mem.h> +#include <vppinfra/format.h> +#include <vppinfra/cache.h> +#include <vlibmemory/unix_shared_memory_queue.h> +#include <signal.h> + +/* + * unix_shared_memory_queue_init + * + * nels = number of elements on the queue + * elsize = element size, presumably 4 and cacheline-size will + * be popular choices. + * coid = consumer coid, from ChannelCreate + * pid = consumer pid + * pulse_code = pulse code consumer expects + * pulse_value = pulse value consumer expects + * consumer_prio = consumer's priority, so pulses won't change + * the consumer's priority. + * + * The idea is to call this function in the queue consumer, + * and e-mail the queue pointer to the producer(s). + * + * The spp process / main thread allocates one of these + * at startup; its main input queue. The spp main input queue + * has a pointer to it in the shared memory segment header. + * + * You probably want to be on an svm data heap before calling this + * function. + */ +unix_shared_memory_queue_t * +unix_shared_memory_queue_init (int nels, + int elsize, + int consumer_pid, + int signal_when_queue_non_empty) +{ + unix_shared_memory_queue_t *q; + pthread_mutexattr_t attr; + pthread_condattr_t cattr; + + q = clib_mem_alloc_aligned (sizeof (unix_shared_memory_queue_t) + + nels * elsize, CLIB_CACHE_LINE_BYTES); + memset (q, 0, sizeof (*q)); + + q->elsize = elsize; + q->maxsize = nels; + q->consumer_pid = consumer_pid; + q->signal_when_queue_non_empty = signal_when_queue_non_empty; + + memset (&attr, 0, sizeof (attr)); + memset (&cattr, 0, sizeof (attr)); + + if (pthread_mutexattr_init (&attr)) + clib_unix_warning ("mutexattr_init"); + if (pthread_mutexattr_setpshared (&attr, PTHREAD_PROCESS_SHARED)) + clib_unix_warning ("pthread_mutexattr_setpshared"); + if (pthread_mutex_init (&q->mutex, &attr)) + clib_unix_warning ("mutex_init"); + if (pthread_mutexattr_destroy (&attr)) + clib_unix_warning ("mutexattr_destroy"); + if (pthread_condattr_init (&cattr)) + clib_unix_warning ("condattr_init"); + /* prints funny-looking messages in the Linux target */ + if (pthread_condattr_setpshared (&cattr, PTHREAD_PROCESS_SHARED)) + clib_unix_warning ("condattr_setpshared"); + if (pthread_cond_init (&q->condvar, &cattr)) + clib_unix_warning ("cond_init1"); + if (pthread_condattr_destroy (&cattr)) + clib_unix_warning ("cond_init2"); + + return (q); +} + +/* + * unix_shared_memory_queue_free + */ +void +unix_shared_memory_queue_free (unix_shared_memory_queue_t * q) +{ + (void) pthread_mutex_destroy (&q->mutex); + (void) pthread_cond_destroy (&q->condvar); + clib_mem_free (q); +} + +void +unix_shared_memory_queue_lock (unix_shared_memory_queue_t * q) +{ + pthread_mutex_lock (&q->mutex); +} + +void +unix_shared_memory_queue_unlock (unix_shared_memory_queue_t * q) +{ + pthread_mutex_unlock (&q->mutex); +} + +int +unix_shared_memory_queue_is_full (unix_shared_memory_queue_t * q) +{ + return q->cursize == q->maxsize; +} + +/* + * unix_shared_memory_queue_add_nolock + */ +int +unix_shared_memory_queue_add_nolock (unix_shared_memory_queue_t * q, + u8 * elem) +{ + i8 *tailp; + int need_broadcast = 0; + + if (PREDICT_FALSE (q->cursize == q->maxsize)) + { + while (q->cursize == q->maxsize) + { + (void) pthread_cond_wait (&q->condvar, &q->mutex); + } + } + + tailp = (i8 *) (&q->data[0] + q->elsize * q->tail); + clib_memcpy (tailp, elem, q->elsize); + + q->tail++; + q->cursize++; + + need_broadcast = (q->cursize == 1); + + if (q->tail == q->maxsize) + q->tail = 0; + + if (need_broadcast) + { + (void) pthread_cond_broadcast (&q->condvar); + if (q->signal_when_queue_non_empty) + kill (q->consumer_pid, q->signal_when_queue_non_empty); + } + return 0; +} + +int +unix_shared_memory_queue_add_raw (unix_shared_memory_queue_t * q, u8 * elem) +{ + i8 *tailp; + + if (PREDICT_FALSE (q->cursize == q->maxsize)) + { + while (q->cursize == q->maxsize) + ; + } + + tailp = (i8 *) (&q->data[0] + q->elsize * q->tail); + clib_memcpy (tailp, elem, q->elsize); + + q->tail++; + q->cursize++; + + if (q->tail == q->maxsize) + q->tail = 0; + return 0; +} + + +/* + * unix_shared_memory_queue_add + */ +int +unix_shared_memory_queue_add (unix_shared_memory_queue_t * q, + u8 * elem, int nowait) +{ + i8 *tailp; + int need_broadcast = 0; + + if (nowait) + { + /* zero on success */ + if (pthread_mutex_trylock (&q->mutex)) + { + return (-1); + } + } + else + pthread_mutex_lock (&q->mutex); + + if (PREDICT_FALSE (q->cursize == q->maxsize)) + { + if (nowait) + { + pthread_mutex_unlock (&q->mutex); + return (-2); + } + while (q->cursize == q->maxsize) + { + (void) pthread_cond_wait (&q->condvar, &q->mutex); + } + } + + tailp = (i8 *) (&q->data[0] + q->elsize * q->tail); + clib_memcpy (tailp, elem, q->elsize); + + q->tail++; + q->cursize++; + + need_broadcast = (q->cursize == 1); + + if (q->tail == q->maxsize) + q->tail = 0; + + if (need_broadcast) + { + (void) pthread_cond_broadcast (&q->condvar); + if (q->signal_when_queue_non_empty) + kill (q->consumer_pid, q->signal_when_queue_non_empty); + } + pthread_mutex_unlock (&q->mutex); + + return 0; +} + +/* + * unix_shared_memory_queue_sub + */ +int +unix_shared_memory_queue_sub (unix_shared_memory_queue_t * q, + u8 * elem, int nowait) +{ + i8 *headp; + int need_broadcast = 0; + + if (nowait) + { + /* zero on success */ + if (pthread_mutex_trylock (&q->mutex)) + { + return (-1); + } + } + else + pthread_mutex_lock (&q->mutex); + + if (PREDICT_FALSE (q->cursize == 0)) + { + if (nowait) + { + pthread_mutex_unlock (&q->mutex); + return (-2); + } + while (q->cursize == 0) + { + (void) pthread_cond_wait (&q->condvar, &q->mutex); + } + } + + headp = (i8 *) (&q->data[0] + q->elsize * q->head); + clib_memcpy (elem, headp, q->elsize); + + q->head++; + if (q->cursize == q->maxsize) + need_broadcast = 1; + + q->cursize--; + + if (q->head == q->maxsize) + q->head = 0; + + if (need_broadcast) + (void) pthread_cond_broadcast (&q->condvar); + + pthread_mutex_unlock (&q->mutex); + + return 0; +} + +int +unix_shared_memory_queue_sub_raw (unix_shared_memory_queue_t * q, u8 * elem) +{ + i8 *headp; + + if (PREDICT_FALSE (q->cursize == 0)) + { + while (q->cursize == 0) + ; + } + + headp = (i8 *) (&q->data[0] + q->elsize * q->head); + clib_memcpy (elem, headp, q->elsize); + + q->head++; + q->cursize--; + + if (q->head == q->maxsize) + q->head = 0; + return 0; +} + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/unix_shared_memory_queue.h b/src/vlibmemory/unix_shared_memory_queue.h new file mode 100644 index 00000000000..f758f17c856 --- /dev/null +++ b/src/vlibmemory/unix_shared_memory_queue.h @@ -0,0 +1,69 @@ +/* + *------------------------------------------------------------------ + * unix_shared_memory_queue.h - shared-memory queues + * + * Copyright (c) 2009 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#ifndef included_unix_shared_memory_queue_h +#define included_unix_shared_memory_queue_h + +#include <pthread.h> +#include <vppinfra/mem.h> + +typedef struct _unix_shared_memory_queue +{ + pthread_mutex_t mutex; /* 8 bytes */ + pthread_cond_t condvar; /* 8 bytes */ + int head; + int tail; + int cursize; + int maxsize; + int elsize; + int consumer_pid; + int signal_when_queue_non_empty; + char data[0]; +} unix_shared_memory_queue_t; + +unix_shared_memory_queue_t *unix_shared_memory_queue_init (int nels, + int elsize, + int consumer_pid, + int + signal_when_queue_non_empty); +void unix_shared_memory_queue_free (unix_shared_memory_queue_t * q); +int unix_shared_memory_queue_add (unix_shared_memory_queue_t * q, + u8 * elem, int nowait); +int unix_shared_memory_queue_sub (unix_shared_memory_queue_t * q, + u8 * elem, int nowait); +void unix_shared_memory_queue_lock (unix_shared_memory_queue_t * q); +void unix_shared_memory_queue_unlock (unix_shared_memory_queue_t * q); +int unix_shared_memory_queue_is_full (unix_shared_memory_queue_t * q); +int unix_shared_memory_queue_add_nolock (unix_shared_memory_queue_t * q, + u8 * elem); + +int unix_shared_memory_queue_sub_raw (unix_shared_memory_queue_t * q, + u8 * elem); +int unix_shared_memory_queue_add_raw (unix_shared_memory_queue_t * q, + u8 * elem); + +#endif /* included_unix_shared_memory_queue_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/vl_memory_api_h.h b/src/vlibmemory/vl_memory_api_h.h new file mode 100644 index 00000000000..c1ae79b1737 --- /dev/null +++ b/src/vlibmemory/vl_memory_api_h.h @@ -0,0 +1,32 @@ +/* + *------------------------------------------------------------------ + * vl_memory_api_h.h - memory API headers, in a specific order. + * + * Copyright (c) 2009-2010 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +/* + * Add to the bottom of the #include list, or elves will steal your + * keyboard in the middle of the night! + */ +#include <vlibmemory/memclnt.api.h> + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vlibmemory/vl_memory_msg_enum.h b/src/vlibmemory/vl_memory_msg_enum.h new file mode 100644 index 00000000000..974c0c21409 --- /dev/null +++ b/src/vlibmemory/vl_memory_msg_enum.h @@ -0,0 +1,42 @@ +/* + *------------------------------------------------------------------ + * vl_memory_msg_enum.h - Our view of how to number API messages + * Clients have their own view, which has to agree with ours. + * + * Copyright (c) 2009-2010 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + *------------------------------------------------------------------ + */ + +#ifndef __VL_MSG_ENUM_H__ +#define __VL_MSG_ENUM_H__ + +#include <vppinfra/byte_order.h> + +#define vl_msg_id(n,h) n, +typedef enum +{ + VL_ILLEGAL_MESSAGE_ID = 0, +#include <vlibmemory/vl_memory_api_h.h> +} vl_msg_id_t; +#undef vl_msg_id + +#endif /* __VL_MSG_ENUM_H__ */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ |