/* * Copyright (c) 2019 Cisco and/or its affiliates. * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this * 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 #include typedef struct vls_shared_data_ { clib_spinlock_t lock; u32 owner_wrk_index; u32 *workers_subscribed; clib_bitmap_t *listeners; } vls_shared_data_t; typedef struct vcl_locked_session_ { clib_spinlock_t lock; u32 session_index; u32 worker_index; u32 vls_index; u32 shared_data_index; /** VCL session owned by different workers because of migration */ u32 owner_vcl_wrk_index; uword *vcl_wrk_index_to_session_index; } vcl_locked_session_t; typedef struct vls_worker_ { vcl_locked_session_t *vls_pool; uword *session_handle_to_vlsh_table; u32 wrk_index; } vls_worker_t; typedef struct vls_local_ { int vls_wrk_index; volatile int vls_mt_n_threads; pthread_mutex_t vls_mt_mq_mlock; pthread_mutex_t vls_mt_spool_mlock; volatile u8 select_mp_check; volatile u8 epoll_mp_check; } vls_process_local_t; static vls_process_local_t vls_local; static vls_process_local_t *vlsl = &vls_local; typedef struct vls_main_ { vls_worker_t *workers; clib_rwlock_t vls_table_lock; /** Pool of data shared by sessions owned by different workers */ vls_shared_data_t *shared_data_pool; clib_rwlock_t shared_data_lock; } vls_main_t; vls_main_t *vlsm; typedef enum vls_rpc_msg_type_ { VLS_RPC_CLONE_AND_SHARE, VLS_RPC_SESS_CLEANUP, } vls_rpc_msg_type_e; typedef struct vls_rpc_msg_ { u8 type; u8 data[0]; } vls_rpc_msg_t; typedef struct vls_clone_and_share_msg_ { u32 vls_index; /**< vls to be shared */ u32 session_index; /**< vcl session to be shared */ u32 origin_vls_wrk; /**< vls worker that initiated the rpc */ u32 origin_vls_index; /**< vls session of the originator */ u32 origin_vcl_wrk; /**< vcl worker that initiated the rpc */ u32 origin_session_index; /**< vcl session of the originator */ } vls_clone_and_share_msg_t; typedef struct vls_sess_cleanup_msg_ { u32 session_index; /**< vcl session to be cleaned */ u32 origin_vcl_wrk; /**< worker that initiated the rpc */ } vls_sess_cleanup_msg_t; void vls_send_session_cleanup_rpc (vcl_worker_t * wrk, u32 dst_wrk_index, u32 dst_session_index); void vls_send_clone_and_share_rpc (vcl_worker_t * wrk, vcl_locked_session_t * vls, u32 session_index, u32 vls_wrk_index, u32 dst_wrk_index, u32 dst_vls_index, u32 dst_session_index); static inline u32 vls_get_worker_index (void) { if (vls_mt_wrk_supported ()) return vlsl->vls_wrk_index; else return vcl_get_worker_index (); } static u32 vls_shared_data_alloc (void) { vls_shared_data_t *vls_shd; u32 shd_index; clib_rwlock_writer_lock (&vlsm->shared_data_lock); pool_get_zero (vlsm->shared_data_pool, vls_shd); clib_spinlock_init (&vls_shd->lock); shd_index = vls_shd - vlsm->shared_data_pool; clib_rwlock_writer_unlock (&vlsm->shared_data_lock); return shd_index; } static u32 vls_shared_data_index (vls_shared_data_t * vls_shd) { return vls_shd - vlsm->shared_data_pool; } vls_shared_data_t * vls_shared_data_get (u32 shd_index) { if (pool_is_free_index (vlsm->shared_data_pool, shd_index)) return 0; return pool_elt_at_index (vlsm->shared_data_pool, shd_index); } static void vls_shared_data_free (u32 shd_index) { vls_shared_data_t *vls_shd; clib_rwlock_writer_lock (&vlsm->shared_data_lock); vls_shd = vls_shared_data_get (shd_index); clib_spinlock_free (&vls_shd->lock); clib_bitmap_free (vls_shd->listeners); vec_free (vls_shd->workers_subscribed); pool_put (vlsm->shared_data_pool, vls_shd); clib_rwlock_writer_unlock (&vlsm->shared_data_lock); } static inline void vls_shared_data_pool_rlock (void) { clib_rwlock_reader_lock (&vlsm->shared_data_lock); } static inline void vls_shared_data_pool_runlock (void) { clib_rwlock_reader_unlock (&vlsm->shared_data_lock); } static inline void vls_mt_table_rlock (void) { if (vlsl->vls_mt_n_threads > 1) clib_rwlock_reader_lock (&vlsm->vls_table_lock); } static inline void vls_mt_table_runlock (void) { if (vlsl->vls_mt_n_threads > 1) clib_rwlock_reader_unlock (&vlsm->vls_table_lock); } static inline void vls_mt_table_wlock (void) { if (vlsl->vls_mt_n_threads > 1) clib_rwlock_writer_lock (&vlsm->vls_table_lock); } static inline void vls_mt_table_wunlock (void) { if (vlsl->vls_mt_n_threads > 1) clib_rwlock_writer_unlock (&vlsm->vls_table_lock); } typedef enum { VLS_MT_OP_READ, VLS_MT_OP_WRITE, VLS_MT_OP_SPOOL, VLS_MT_OP_XPOLL, } vls_mt_ops_t; typedef enum { VLS_MT_LOCK_MQ = 1 << 0, VLS_MT_LOCK_SPOOL = 1 << 1 } vls_mt_lock_type_t; static void vls_mt_add (void) { vlsl->vls_mt_n_threads += 1; /* If multi-thread workers are supported, for each new thread register a new * vcl worker with vpp. Otherwise, all threads use the same vcl worker, so * update the vcl worker's thread local worker index variable */ if (vls_mt_wrk_supported ()) vls_register_vcl_worker (); else vcl_set_worker_index (vlsl->vls_wrk_index); } static inline void vls_mt_mq_lock (void) { pthread_mutex_lock (&vlsl->vls_mt_mq_mlock); } static inline void vls_mt_mq_unlock (void) { pthread_mutex_unlock (&vlsl->vls_mt_mq_mlock); } static inline void vls_mt_spool_lock (void) { pthread_mutex_lock (&vlsl->vls_mt_spool_mlock); } static inline void vls_mt_create_unlock (void) { pthread_mutex_unlock (&vlsl->vls_mt_spool_mlock); } static void vls_mt_locks_init (void) { pthread_mutex_init (&vlsl->vls_mt_mq_mlock, NULL); pthread_mutex_init (&vlsl->vls_mt_spool_mlock, NULL); } u8 vls_is_shared (vcl_locked_session_t * vls) { return (vls->shared_data_index != ~0); } static inline void vls_lock (vcl_locked_session_t * vls) { if ((vlsl->vls_mt_n_threads > 1) || vls_is_shared (vls)) clib_spinlock_lock (&vls->lock); } static inline void vls_unlock (vcl_locked_session_t * vls) { if ((vlsl->vls_mt_n_threads > 1) || vls_is_shared (vls)) clib_spinlock_unlock (&vls->lock); } static inline vcl_session_handle_t vls_to_sh (vcl_locked_session_t * vls) { return vcl_session_handle_from_index (vls->session_index); } static inline vcl_session_handle_t vls_to_sh_tu (vcl_locked_session_t * vls) { vcl_session_handle_t sh; sh = vls_to_sh (vls); vls_mt_table_runlock (); return sh; } static vls_worker_t * vls_worker_get_current (void) { return pool_elt_at_index (vlsm->workers, vls_get_worker_index ()); } static void vls_worker_alloc (void) { vls_worker_t *wrk; pool_get_zero (vlsm->workers, wrk); wrk->wrk_index = vcl_get_worker_index (); } static void vls_worker_free (vls_worker_t * wrk) { hash_free (wrk->session_handle_to_vlsh_table); pool_free (wrk->vls_pool); pool_put (vlsm->workers, wrk); } static vls_worker_t * vls_worker_get (u32 wrk_index) { if (pool_is_free_index (vlsm->workers, wrk_index)) return 0; return pool_elt_at_index (vlsm->workers, wrk_index); } static vls_handle_t vls_alloc (vcl_session_handle_t sh) { vls_worker_t *
/*
 * Copyright (c) 2018 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 __GBP_INTERFACE_H__
#define __GBP_INTERFACE_H__

#include <vnet/l2/l2_input.h>
#include <vnet/l2/l2_output.h>


#define foreach_gdb_l3_feature                  \
  _(LEARN_IP4, "gbp-learn-ip4", "ip4-unicast") \
  _(LEARN_IP6, "gbp-learn-ip6", "ip6-unicast")

typedef enum gbp_itf_l3_feat_pos_t_
{
#define _(s,v,a) GBP_ITF_L3_FEAT_POS_##s,
  foreach_gdb_l3_feature
#undef _
} gbp_itf_l3_feat_pos_t;

typedef enum gbp_itf_l3_feat_t_
{
  GBP_ITF_L3_FEAT_NONE,
#define _(s,v,a) GBP_ITF_L3_FEAT_##s = (1 << GBP_ITF_L3_FEAT_POS_##s),
  foreach_gdb_l3_feature
#undef _
} gbp_itf_l3_feat_t;

#define GBP_ITF_L3_FEAT_LEARN (GBP_ITF_L3_FEAT_LEARN_IP4|GBP_ITF_L3_FEAT_LEARN_IP6)

typedef struct gbp_itf_hdl_t_
{
  union
  {
    struct
    {
      u32 gh_who;
      u32 gh_which;
    };
  };
} gbp_itf_hdl_t;

#define GBP_ITF_HDL_INIT {.gh_which = ~0}
const static gbp_itf_hdl_t GBP_ITF_HDL_INVALID = GBP_ITF_HDL_INIT;

extern void gbp_itf_hdl_reset (gbp_itf_hdl_t * gh);
extern bool gbp_itf_hdl_is_valid (gbp_itf_hdl_t gh);

typedef void (*gbp_itf_free_fn_t) (u32 sw_if_index);

extern gbp_itf_hdl_t gbp_itf_l2_add_and_lock (u32 sw_if_index, u32 bd_index);
extern gbp_itf_hdl_t gbp_itf_l3_add_and_lock (u32 sw_if_index, index_t gri);
extern gbp_itf_hdl_t gbp_itf_l2_add_and_lock_w_free (u32 sw_if_index,
						     u32 bd_index,
						     gbp_itf_free_fn_t ff);
extern gbp_itf_hdl_t gbp_itf_l3_add_and_lock_w_free (u32 sw_if_index,
						     index_t gri,
						     gbp_itf_free_fn_t ff);

extern void gbp_itf_unlock (gbp_itf_hdl_t * hdl);
extern void gbp_itf_lock (gbp_itf_hdl_t hdl);
extern gbp_itf_hdl_t gbp_itf_clone_and_lock (gbp_itf_hdl_t hdl);
extern u32 gbp_itf_get_sw_if_index (gbp_itf_hdl_t hdl);

extern void gbp_itf_l2_set_input_feature (gbp_itf_hdl_t hdl,
					  l2input_feat_masks_t feats);
extern void gbp_itf_l2_set_output_feature (gbp_itf_hdl_t hdl,
					   l2output_feat_masks_t feats);

extern void gbp_itf_l3_set_input_feature (gbp_itf_hdl_t hdl,
					  gbp_itf_l3_feat_t feats);

extern u8 *format_gbp_itf_hdl (u8 * s, va_list * args);

#endif

/*
 * fd.io coding-style-patch-verification: ON
 *
 * Local Variables:
 * eval: (c-set-style "gnu")
 * End:
 */
n of worker (session): %u (%u) -> %u (%u)", vls->owner_vcl_wrk_index, src_sid, wrk_index, sid); if (PREDICT_FALSE (session->is_vep && session->vep.next_sh != ~0)) { /* TODO: rollback? */ VERR ("can't migrate nonempty epoll session"); ASSERT (0); return; } else if (PREDICT_FALSE (!session->is_vep && session->session_state != STATE_CLOSED)) { /* TODO: rollback? */ VERR ("migrate NOT supported, session_status (%u)", session->session_state); ASSERT (0); return; } } static inline void vls_mt_detect (void) { if (PREDICT_FALSE (vcl_get_worker_index () == ~0)) vls_mt_add (); } #define vls_mt_guard(_vls, _op) \ int _locks_acq = 0; \ if (vls_mt_wrk_supported ()) \ { \ if (PREDICT_FALSE (_vls \ && ((vcl_locked_session_t *)_vls)->worker_index != \ vcl_get_worker_index ())) \ vls_mt_session_migrate (_vls); \ } \ else \ { \ if (PREDICT_FALSE (vlsl->vls_mt_n_threads > 1)) \ vls_mt_acq_locks (_vls, _op, &_locks_acq); \ } \ #define vls_mt_unguard() \ if (PREDICT_FALSE (_locks_acq)) \ vls_mt_rel_locks (_locks_acq) int vls_write (vls_handle_t vlsh, void *buf, size_t nbytes) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_WRITE); rv = vppcom_session_write (vls_to_sh_tu (vls), buf, nbytes); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } int vls_write_msg (vls_handle_t vlsh, void *buf, size_t nbytes) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_WRITE); rv = vppcom_session_write_msg (vls_to_sh_tu (vls), buf, nbytes); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } int vls_sendto (vls_handle_t vlsh, void *buf, int buflen, int flags, vppcom_endpt_t * ep) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_WRITE); rv = vppcom_session_sendto (vls_to_sh_tu (vls), buf, buflen, flags, ep); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } ssize_t vls_read (vls_handle_t vlsh, void *buf, size_t nbytes) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_READ); rv = vppcom_session_read (vls_to_sh_tu (vls), buf, nbytes); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } ssize_t vls_recvfrom (vls_handle_t vlsh, void *buffer, uint32_t buflen, int flags, vppcom_endpt_t * ep) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_READ); rv = vppcom_session_recvfrom (vls_to_sh_tu (vls), buffer, buflen, flags, ep); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } int vls_attr (vls_handle_t vlsh, uint32_t op, void *buffer, uint32_t * buflen) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; if (vls_mt_session_should_migrate (vls)) vls_mt_session_migrate (vls); rv = vppcom_session_attr (vls_to_sh_tu (vls), op, buffer, buflen); vls_get_and_unlock (vlsh); return rv; } int vls_bind (vls_handle_t vlsh, vppcom_endpt_t * ep) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; rv = vppcom_session_bind (vls_to_sh_tu (vls), ep); vls_get_and_unlock (vlsh); return rv; } int vls_listen (vls_handle_t vlsh, int q_len) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_XPOLL); rv = vppcom_session_listen (vls_to_sh_tu (vls), q_len); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } int vls_connect (vls_handle_t vlsh, vppcom_endpt_t * server_ep) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (vlsh))) return VPPCOM_EBADFD; vls_mt_guard (vls, VLS_MT_OP_XPOLL); rv = vppcom_session_connect (vls_to_sh_tu (vls), server_ep); vls_mt_unguard (); vls_get_and_unlock (vlsh); return rv; } static inline void vls_mp_checks (vcl_locked_session_t * vls, int is_add) { vcl_worker_t *wrk = vcl_worker_get_current (); vcl_session_t *s; u32 owner_wrk; if (vls_mt_wrk_supported ()) return; s = vcl_session_get (wrk, vls->session_index); switch (s->session_state) { case STATE_LISTEN: if (is_add) { vls_listener_wrk_set (vls, vls->worker_index, 1 /* is_active */ ); break; } vls_listener_wrk_stop_listen (vls, vls->worker_index); break; case STATE_LISTEN_NO_MQ: if (!is_add) break; /* Register worker as listener */ vls_listener_wrk_start_listen (vls, wrk->wrk_index); /* If owner worker did not attempt to accept/xpoll on the session, * force a listen stop for it, since it may not be interested in * accepting new sessions. * This is pretty much a hack done to give app workers the illusion * that it is fine to listen and not accept new sessions for a * given listener. Without it, we would accumulate unhandled * accepts on the passive worker message queue. */ owner_wrk = vls_shared_get_owner (vls); if (!vls_listener_wrk_is_active (vls, owner_wrk)) vls_listener_wrk_stop_listen (vls, owner_wrk); break; default: break; } } vls_handle_t vls_accept (vls_handle_t listener_vlsh, vppcom_endpt_t * ep, int flags) { vls_handle_t accepted_vlsh; vcl_locked_session_t *vls; int sh; vls_mt_detect (); if (!(vls = vls_get_w_dlock (listener_vlsh))) return VPPCOM_EBADFD; if (vcl_n_workers () > 1) vls_mp_checks (vls, 1 /* is_add */ ); vls_mt_guard (vls, VLS_MT_OP_SPOOL); sh = vppcom_session_accept (vls_to_sh_tu (vls), ep, flags); vls_mt_unguard (); vls_get_and_unlock (listener_vlsh); if (sh < 0) return sh; accepted_vlsh = vls_alloc (sh); if (PREDICT_FALSE (accepted_vlsh == VLS_INVALID_HANDLE)) vppcom_session_close (sh); return accepted_vlsh; } vls_handle_t vls_create (uint8_t proto, uint8_t is_nonblocking) { vcl_session_handle_t sh; vls_handle_t vlsh; vls_mt_detect (); vls_mt_guard (0, VLS_MT_OP_SPOOL); sh = vppcom_session_create (proto, is_nonblocking); vls_mt_unguard (); if (sh == INVALID_SESSION_ID) return VLS_INVALID_HANDLE; vlsh = vls_alloc (sh); if (PREDICT_FALSE (vlsh == VLS_INVALID_HANDLE)) vppcom_session_close (sh); return vlsh; } static void vls_mt_session_cleanup (vcl_locked_session_t * vls) { u32 session_index, wrk_index, current_vcl_wrk; vcl_worker_t *wrk = vcl_worker_get_current (); ASSERT (vls_mt_wrk_supported ()); current_vcl_wrk = vcl_get_worker_index (); /* *INDENT-OFF* */ hash_foreach (wrk_index, session_index, vls->vcl_wrk_index_to_session_index, ({ if (current_vcl_wrk != wrk_index) vls_send_session_cleanup_rpc (wrk, wrk_index, session_index); })); /* *INDENT-ON* */ hash_free (vls->vcl_wrk_index_to_session_index); } int vls_close (vls_handle_t vlsh) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); vls_mt_table_wlock (); vls = vls_get_and_lock (vlsh); if (!vls) { vls_mt_table_wunlock (); return VPPCOM_EBADFD; } vls_mt_guard (vls, VLS_MT_OP_SPOOL); if (vls_is_shared (vls)) rv = vls_unshare_session (vls, vcl_worker_get_current ()); else rv = vppcom_session_close (vls_to_sh (vls)); if (vls_mt_wrk_supported ()) vls_mt_session_cleanup (vls); vls_free (vls); vls_mt_unguard (); vls_mt_table_wunlock (); return rv; } vls_handle_t vls_epoll_create (void) { vcl_session_handle_t sh; vls_handle_t vlsh; vls_mt_detect (); sh = vppcom_epoll_create (); if (sh == INVALID_SESSION_ID) return VLS_INVALID_HANDLE; vlsh = vls_alloc (sh); if (vlsh == VLS_INVALID_HANDLE) vppcom_session_close (sh); return vlsh; } static void vls_epoll_ctl_mp_checks (vcl_locked_session_t * vls, int op) { if (vcl_n_workers () <= 1) { vlsl->epoll_mp_check = 1; return; } if (op == EPOLL_CTL_MOD) return; vlsl->epoll_mp_check = 1; vls_mp_checks (vls, op == EPOLL_CTL_ADD); } int vls_epoll_ctl (vls_handle_t ep_vlsh, int op, vls_handle_t vlsh, struct epoll_event *event) { vcl_locked_session_t *ep_vls, *vls; vcl_session_handle_t ep_sh, sh; int rv; vls_mt_detect (); vls_mt_table_rlock (); ep_vls = vls_get_and_lock (ep_vlsh); vls = vls_get_and_lock (vlsh); if (vls_mt_session_should_migrate (ep_vls)) vls_mt_session_migrate (ep_vls); ep_sh = vls_to_sh (ep_vls); sh = vls_to_sh (vls); if (PREDICT_FALSE (!vlsl->epoll_mp_check)) vls_epoll_ctl_mp_checks (vls, op); vls_mt_table_runlock (); rv = vppcom_epoll_ctl (ep_sh, op, sh, event); vls_mt_table_rlock (); ep_vls = vls_get (ep_vlsh); vls = vls_get (vlsh); vls_unlock (vls); vls_unlock (ep_vls); vls_mt_table_runlock (); return rv; } int vls_epoll_wait (vls_handle_t ep_vlsh, struct epoll_event *events, int maxevents, double wait_for_time) { vcl_locked_session_t *vls; int rv; vls_mt_detect (); if (!(vls = vls_get_w_dlock (ep_vlsh))) return VPPCOM_EBADFD; vls_mt_guard (0, VLS_MT_OP_XPOLL); rv = vppcom_epoll_wait (vls_to_sh_tu (vls), events, maxevents, wait_for_time); vls_mt_unguard (); vls_get_and_unlock (ep_vlsh); return rv; } static void vls_select_mp_checks (vcl_si_set * read_map) { vcl_locked_session_t *vls; vcl_worker_t *wrk; vcl_session_t *s; u32 si; if (vcl_n_workers () <= 1) { vlsl->select_mp_check = 1; return; } if (!read_map) return; vlsl->select_mp_check = 1; wrk = vcl_worker_get_current (); /* *INDENT-OFF* */ clib_bitmap_foreach (si, read_map, ({ s = vcl_session_get (wrk, si); if (s->session_state == STATE_LISTEN) { vls = vls_get (vls_session_index_to_vlsh (si)); vls_mp_checks (vls, 1 /* is_add */); } })); /* *INDENT-ON* */ } int vls_select (int n_bits, vcl_si_set * read_map, vcl_si_set * write_map, vcl_si_set * except_map, double wait_for_time) { int rv; vls_mt_detect (); vls_mt_guard (0, VLS_MT_OP_XPOLL); if (PREDICT_FALSE (!vlsl->select_mp_check)) vls_select_mp_checks (read_map); rv = vppcom_select (n_bits, read_map, write_map, except_map, wait_for_time); vls_mt_unguard (); return rv; } static void vls_unshare_vcl_worker_sessions (vcl_worker_t * wrk) { u32 current_wrk, is_current; vcl_locked_session_t *vls; vcl_session_t *s; if (pool_elts (vcm->workers) <= 1) return; current_wrk = vcl_get_worker_index (); is_current = current_wrk == wrk->wrk_index; /* *INDENT-OFF* */ pool_foreach (s, wrk->sessions, ({ vls = vls_get (vls_si_wi_to_vlsh (s->session_index, wrk->wrk_index)); if (vls && (is_current || vls_is_shared_by_wrk (vls, current_wrk))) vls_unshare_session (vls, wrk); })); /* *INDENT-ON* */ } static void vls_cleanup_vcl_worker (vcl_worker_t * wrk) { vls_worker_t *vls_wrk = vls_worker_get (wrk->wrk_index); /* Unshare sessions and also cleanup worker since child may have * called _exit () and therefore vcl may not catch the event */ vls_unshare_vcl_worker_sessions (wrk); vcl_worker_cleanup (wrk, 1 /* notify vpp */ ); vls_worker_free (vls_wrk); } static void vls_cleanup_forked_child (vcl_worker_t * wrk, vcl_worker_t * child_wrk) { vcl_worker_t *sub_child; int tries = 0; if (child_wrk->forked_child != ~0) { sub_child = vcl_worker_get_if_valid (child_wrk->forked_child); if (sub_child) { /* Wait a bit, maybe the process is going away */ while (kill (sub_child->current_pid, 0) >= 0 && tries++ < 50) usleep (1e3); if (kill (sub_child->current_pid, 0) < 0) vls_cleanup_forked_child (child_wrk, sub_child); } } vls_cleanup_vcl_worker (child_wrk); VDBG (0, "Cleaned up forked child wrk %u", child_wrk->wrk_index); wrk->forked_child = ~0; } static struct sigaction old_sa; static void vls_intercept_sigchld_handler (int signum, siginfo_t * si, void *uc) { vcl_worker_t *wrk, *child_wrk; if (vcl_get_worker_index () == ~0) return; if (sigaction (SIGCHLD, &old_sa, 0)) { VERR ("couldn't restore sigchld"); exit (-1); } wrk = vcl_worker_get_current (); if (wrk->forked_child == ~0) return; child_wrk = vcl_worker_get_if_valid (wrk->forked_child); if (!child_wrk) goto done; if (si && si->si_pid != child_wrk->current_pid) { VDBG (0, "unexpected child pid %u", si->si_pid); goto done; } vls_cleanup_forked_child (wrk, child_wrk); done: if (old_sa.sa_flags & SA_SIGINFO) { void (*fn) (int, siginfo_t *, void *) = old_sa.sa_sigaction; fn (signum, si, uc); } else { void (*fn) (int) = old_sa.sa_handler; if (fn) fn (signum); } } static void vls_incercept_sigchld () { struct sigaction sa; clib_memset (&sa, 0, sizeof (sa)); sa.sa_sigaction = vls_intercept_sigchld_handler; sa.sa_flags = SA_SIGINFO; if (sigaction (SIGCHLD, &sa, &old_sa)) { VERR ("couldn't intercept sigchld"); exit (-1); } } static void vls_app_pre_fork (void) { vls_incercept_sigchld (); vcl_flush_mq_events (); } static void vls_app_fork_child_handler (void) { vcl_worker_t *parent_wrk; int parent_wrk_index; parent_wrk_index = vcl_get_worker_index (); VDBG (0, "initializing forked child %u with parent wrk %u", getpid (), parent_wrk_index); /* * Clear old state */ vcl_set_worker_index (~0); /* * Allocate and register vcl worker with vpp */ if (vppcom_worker_register ()) { VERR ("couldn't register new worker!"); return; } /* * Allocate/initialize vls worker and share sessions */ vls_worker_alloc (); parent_wrk = vcl_worker_get (parent_wrk_index); vls_worker_copy_on_fork (parent_wrk); parent_wrk->forked_child = vcl_get_worker_index (); /* Reset number of threads and set wrk index */ vlsl->vls_mt_n_threads = 0; vlsl->vls_wrk_index = vcl_get_worker_index (); vlsl->select_mp_check = 0; vlsl->epoll_mp_check = 0; vls_mt_locks_init (); VDBG (0, "forked child main worker initialized"); vcm->forking = 0; } static void vls_app_fork_parent_handler (void) { vcm->forking = 1; while (vcm->forking) ; } void vls_app_exit (void) { vls_worker_t *wrk = vls_worker_get_current (); /* Unshare the sessions. VCL will clean up the worker */ vls_unshare_vcl_worker_sessions (vcl_worker_get_current ()); vls_worker_free (wrk); } static void vls_clone_and_share_rpc_handler (void *args) { vls_clone_and_share_msg_t *msg = (vls_clone_and_share_msg_t *) args; vls_worker_t *wrk = vls_worker_get_current (), *dst_wrk; vcl_locked_session_t *vls, *dst_vls; vcl_worker_t *vcl_wrk = vcl_worker_get_current (), *dst_vcl_wrk; vcl_session_t *s, *dst_s; vls = vls_session_get (wrk, msg->vls_index); if (!vls_mt_wrk_supported ()) vls_init_share_session (wrk, vls); s = vcl_session_get (vcl_wrk, msg->session_index); dst_wrk = vls_worker_get (msg->origin_vls_wrk); dst_vcl_wrk = vcl_worker_get (msg->origin_vcl_wrk); dst_vls = vls_session_get (dst_wrk, msg->origin_vls_index); dst_vls->shared_data_index = vls->shared_data_index; dst_s = vcl_session_get (dst_vcl_wrk, msg->origin_session_index); clib_memcpy (dst_s, s, sizeof (*s)); dst_vcl_wrk->rpc_done = 1; VDBG (1, "proces session clone of worker (session): %u (%u) -> %u (%u)", vcl_wrk->wrk_index, msg->session_index, dst_vcl_wrk->wrk_index, msg->origin_session_index); } static void vls_session_cleanup_rpc_handler (void *args) { vls_sess_cleanup_msg_t *msg = (vls_sess_cleanup_msg_t *) args; vcl_worker_t *wrk = vcl_worker_get_current (); vcl_worker_t *dst_wrk = vcl_worker_get (msg->origin_vcl_wrk); vppcom_session_close (vcl_session_handle_from_index (msg->session_index)); dst_wrk->rpc_done = 1; VDBG (1, "proces session cleanup of worker (session): %u (%u) from %u ()", wrk->wrk_index, msg->session_index, dst_wrk->wrk_index); } static void vls_rpc_handler (void *args) { vls_rpc_msg_t *msg = (vls_rpc_msg_t *) args; switch (msg->type) { case VLS_RPC_CLONE_AND_SHARE: vls_clone_and_share_rpc_handler (msg->data); break; case VLS_RPC_SESS_CLEANUP: vls_session_cleanup_rpc_handler (msg->data); break; default: break; } } void vls_send_clone_and_share_rpc (vcl_worker_t * wrk, vcl_locked_session_t * vls, u32 session_index, u32 vls_wrk_index, u32 dst_wrk_index, u32 dst_vls_index, u32 dst_session_index) { u8 data[sizeof (u8) + sizeof (vls_clone_and_share_msg_t)]; vls_clone_and_share_msg_t *msg; vls_rpc_msg_t *rpc; int ret; rpc = (vls_rpc_msg_t *) & data; rpc->type = VLS_RPC_CLONE_AND_SHARE; msg = (vls_clone_and_share_msg_t *) & rpc->data; msg->origin_vls_wrk = vls_wrk_index; msg->origin_vls_index = vls->vls_index; msg->origin_vcl_wrk = wrk->wrk_index; msg->origin_session_index = session_index; msg->vls_index = dst_vls_index; msg->session_index = dst_session_index; wrk->rpc_done = 0; ret = vcl_send_worker_rpc (dst_wrk_index, rpc, sizeof (data)); VDBG (1, "send session clone to wrk (session): %u (%u) -> %u (%u), ret=%d", dst_wrk_index, msg->session_index, msg->origin_vcl_wrk, msg->origin_session_index, ret); while (!ret && !wrk->rpc_done) ; } void vls_send_session_cleanup_rpc (vcl_worker_t * wrk, u32 dst_wrk_index, u32 dst_session_index) { u8 data[sizeof (u8) + sizeof (vls_sess_cleanup_msg_t)]; vls_sess_cleanup_msg_t *msg; vls_rpc_msg_t *rpc; int ret; rpc = (vls_rpc_msg_t *) & data; rpc->type = VLS_RPC_SESS_CLEANUP; msg = (vls_sess_cleanup_msg_t *) & rpc->data; msg->origin_vcl_wrk = wrk->wrk_index; msg->session_index = dst_session_index; wrk->rpc_done = 0; ret = vcl_send_worker_rpc (dst_wrk_index, rpc, sizeof (data)); VDBG (1, "send session cleanup to wrk (session): %u (%u) from %u, ret=%d", dst_wrk_index, msg->session_index, msg->origin_vcl_wrk, ret); while (!ret && !wrk->rpc_done) ; } int vls_app_create (char *app_name) { int rv; if ((rv = vppcom_app_create (app_name))) return rv; vlsm = clib_mem_alloc (sizeof (vls_main_t)); clib_memset (vlsm, 0, sizeof (*vlsm)); clib_rwlock_init (&vlsm->vls_table_lock); clib_rwlock_init (&vlsm->shared_data_lock); pool_alloc (vlsm->workers, vcm->cfg.max_workers); pthread_atfork (vls_app_pre_fork, vls_app_fork_parent_handler, vls_app_fork_child_handler); atexit (vls_app_exit); vls_worker_alloc (); vlsl->vls_wrk_index = vcl_get_worker_index (); vls_mt_locks_init (); vcm->wrk_rpc_fn = vls_rpc_handler; return VPPCOM_OK; } unsigned char vls_use_eventfd (void) { return vcm->cfg.use_mq_eventfd; } unsigned char vls_mt_wrk_supported (void) { return vcm->cfg.mt_wrk_supported; } int vls_use_real_epoll (void) { if (vcl_get_worker_index () == ~0) return 0; return vcl_worker_get_current ()->vcl_needs_real_epoll; } void vls_register_vcl_worker (void) { if (vppcom_worker_register () != VPPCOM_OK) { VERR ("failed to register worker"); return; } } /* * fd.io coding-style-patch-verification: ON * * Local Variables: * eval: (c-set-style "gnu") * End: */