diff --git a/libknet/internals.h b/libknet/internals.h index c1c71509..70e635fe 100644 --- a/libknet/internals.h +++ b/libknet/internals.h @@ -1,504 +1,506 @@ /* * Copyright (C) 2010-2017 Red Hat, Inc. All rights reserved. * * Authors: Fabio M. Di Nitto * Federico Simoncelli * * This software licensed under GPL-2.0+, LGPL-2.0+ */ #ifndef __KNET_INTERNALS_H__ #define __KNET_INTERNALS_H__ /* * NOTE: you shouldn't need to include this header normally */ #include #include "libknet.h" #include "onwire.h" #include "compat.h" #define KNET_DATABUFSIZE KNET_MAX_PACKET_SIZE + KNET_HEADER_ALL_SIZE #define KNET_DATABUFSIZE_CRYPT_PAD 1024 #define KNET_DATABUFSIZE_CRYPT KNET_DATABUFSIZE + KNET_DATABUFSIZE_CRYPT_PAD #define KNET_DATABUFSIZE_COMPRESS_PAD 1024 #define KNET_DATABUFSIZE_COMPRESS KNET_DATABUFSIZE + KNET_DATABUFSIZE_COMPRESS_PAD #define KNET_RING_RCVBUFF 8388608 #define PCKT_FRAG_MAX UINT8_MAX #define PCKT_RX_BUFS 512 #define KNET_EPOLL_MAX_EVENTS KNET_DATAFD_MAX typedef void *knet_transport_link_t; /* per link transport handle */ typedef void *knet_transport_t; /* per knet_h transport handle */ struct knet_transport_ops; /* Forward because of circular dependancy */ struct knet_mmsghdr { struct msghdr msg_hdr; /* Message header */ unsigned int msg_len; /* Number of bytes transmitted */ }; struct knet_link { /* required */ struct sockaddr_storage src_addr; struct sockaddr_storage dst_addr; /* configurable */ unsigned int dynamic; /* see KNET_LINK_DYN_ define above */ uint8_t priority; /* higher priority == preferred for A/P */ unsigned long long ping_interval; /* interval */ unsigned long long pong_timeout; /* timeout */ unsigned long long pong_timeout_adj; /* timeout adjusted for latency */ uint8_t pong_timeout_backoff; /* see link.h for definition */ unsigned int latency_fix; /* precision */ uint8_t pong_count; /* how many ping/pong to send/receive before link is up */ uint64_t flags; /* status */ struct knet_link_status status; /* internals */ uint8_t link_id; uint8_t transport_type; /* #defined constant from API */ knet_transport_link_t transport_link; /* link_info_t from transport */ int outsock; unsigned int configured:1; /* set to 1 if src/dst have been configured transport initialized on this link*/ unsigned int transport_connected:1; /* set to 1 if lower level transport is connected */ unsigned int latency_exp; uint8_t received_pong; struct timespec ping_last; /* used by PMTUD thread as temp per-link variables and should always contain the onwire_len value! */ uint32_t proto_overhead; struct timespec pmtud_last; uint32_t last_ping_size; uint32_t last_good_mtu; uint32_t last_bad_mtu; uint32_t last_sent_mtu; uint32_t last_recv_mtu; uint8_t has_valid_mtu; }; #define KNET_CBUFFER_SIZE 4096 struct knet_host_defrag_buf { char buf[KNET_DATABUFSIZE]; uint8_t in_use; /* 0 buffer is free, 1 is in use */ seq_num_t pckt_seq; /* identify the pckt we are receiving */ uint8_t frag_recv; /* how many frags did we receive */ uint8_t frag_map[PCKT_FRAG_MAX];/* bitmap of what we received? */ uint8_t last_first; /* special case if we receive the last fragment first */ uint16_t frag_size; /* normal frag size (not the last one) */ uint16_t last_frag_size; /* the last fragment might not be aligned with MTU size */ struct timespec last_update; /* keep time of the last pckt */ }; struct knet_host { /* required */ knet_node_id_t host_id; /* configurable */ uint8_t link_handler_policy; char name[KNET_MAX_HOST_LEN]; /* status */ struct knet_host_status status; /* internals */ char circular_buffer[KNET_CBUFFER_SIZE]; seq_num_t rx_seq_num; seq_num_t untimed_rx_seq_num; seq_num_t timed_rx_seq_num; uint8_t got_data; /* defrag/reassembly buffers */ struct knet_host_defrag_buf defrag_buf[KNET_MAX_LINK]; char circular_buffer_defrag[KNET_CBUFFER_SIZE]; /* link stuff */ struct knet_link link[KNET_MAX_LINK]; uint8_t active_link_entries; uint8_t active_links[KNET_MAX_LINK]; struct knet_host *next; }; struct knet_sock { int sockfd[2]; /* sockfd[0] will always be application facing * and sockfd[1] internal if sockpair has been created by knet */ int is_socket; /* check if it's a socket for recvmmsg usage */ int is_created; /* knet created this socket and has to clean up on exit/del */ int in_use; /* set to 1 if it's use, 0 if free */ int has_error; /* set to 1 if there were errors reading from the sock * and socket has been removed from epoll */ }; struct knet_fd_trackers { uint8_t transport; /* transport type (UDP/SCTP...) */ uint8_t data_type; /* internal use for transport to define what data are associated * to this fd */ void *data; /* pointer to the data */ }; #define KNET_MAX_FDS KNET_MAX_HOST * KNET_MAX_LINK * 4 #define KNET_MAX_COMPRESS_METHODS UINT8_MAX struct knet_handle_stats_extra { uint64_t tx_crypt_pmtu_packets; uint64_t tx_crypt_pmtu_reply_packets; uint64_t tx_crypt_ping_packets; uint64_t tx_crypt_pong_packets; }; struct knet_handle { knet_node_id_t host_id; unsigned int enabled:1; struct knet_sock sockfd[KNET_DATAFD_MAX]; int logfd; uint8_t log_levels[KNET_MAX_SUBSYSTEMS]; int hostsockfd[2]; int dstsockfd[2]; int send_to_links_epollfd; int recv_from_links_epollfd; int dst_link_handler_epollfd; unsigned int pmtud_interval; unsigned int data_mtu; /* contains the max data size that we can send onwire * without frags */ struct knet_host *host_head; struct knet_host *host_index[KNET_MAX_HOST]; knet_transport_t transports[KNET_MAX_TRANSPORTS+1]; struct knet_fd_trackers knet_transport_fd_tracker[KNET_MAX_FDS]; /* track status for each fd handled by transports */ struct knet_handle_stats stats; struct knet_handle_stats_extra stats_extra; uint32_t reconnect_int; knet_node_id_t host_ids[KNET_MAX_HOST]; size_t host_ids_entries; struct knet_header *recv_from_sock_buf; struct knet_header *send_to_links_buf[PCKT_FRAG_MAX]; struct knet_header *recv_from_links_buf[PCKT_RX_BUFS]; struct knet_header *pingbuf; struct knet_header *pmtudbuf; pthread_t send_to_links_thread; pthread_t recv_from_links_thread; pthread_t heartbt_thread; pthread_t dst_link_handler_thread; pthread_t pmtud_link_handler_thread; int lock_init_done; pthread_rwlock_t global_rwlock; /* global config lock */ pthread_mutex_t pmtud_mutex; /* pmtud mutex to handle conditional send/recv + timeout */ pthread_cond_t pmtud_cond; /* conditional for above */ pthread_mutex_t tx_mutex; /* used to protect knet_send_sync and TX thread */ pthread_mutex_t hb_mutex; /* used to protect heartbeat thread and seq_num broadcasting */ pthread_mutex_t backoff_mutex; /* used to protect dst_link->pong_timeout_adj */ pthread_mutex_t kmtu_mutex; /* used to protect kernel_mtu */ uint32_t kernel_mtu; /* contains the MTU detected by the kernel on a given link */ int pmtud_waiting; + int pmtud_running; + int pmtud_forcerun; int pmtud_abort; struct crypto_instance *crypto_instance; size_t sec_header_size; size_t sec_block_size; size_t sec_hash_size; size_t sec_salt_size; unsigned char *send_to_links_buf_crypt[PCKT_FRAG_MAX]; unsigned char *recv_from_links_buf_crypt; unsigned char *recv_from_links_buf_decrypt; unsigned char *pingbuf_crypt; unsigned char *pmtudbuf_crypt; int compress_model; int compress_level; size_t compress_threshold; void *compress_int_data[KNET_MAX_COMPRESS_METHODS]; /* for compress method private data */ unsigned char *recv_from_links_buf_decompress; unsigned char *send_to_links_buf_compress; seq_num_t tx_seq_num; pthread_mutex_t tx_seq_num_mutex; uint8_t has_loop_link; uint8_t loop_link; void *dst_host_filter_fn_private_data; int (*dst_host_filter_fn) ( void *private_data, const unsigned char *outdata, ssize_t outdata_len, uint8_t tx_rx, knet_node_id_t this_host_id, knet_node_id_t src_node_id, int8_t *channel, knet_node_id_t *dst_host_ids, size_t *dst_host_ids_entries); void *pmtud_notify_fn_private_data; void (*pmtud_notify_fn) ( void *private_data, unsigned int data_mtu); void *host_status_change_notify_fn_private_data; void (*host_status_change_notify_fn) ( void *private_data, knet_node_id_t host_id, uint8_t reachable, uint8_t remote, uint8_t external); void *sock_notify_fn_private_data; void (*sock_notify_fn) ( void *private_data, int datafd, int8_t channel, uint8_t tx_rx, int error, int errorno); int fini_in_progress; }; extern pthread_rwlock_t shlib_rwlock; /* global shared lib load lock */ /* * NOTE: every single operation must be implementend * for every protocol. */ typedef struct knet_transport_ops { /* * transport generic information */ const char *transport_name; const uint8_t transport_id; const uint8_t built_in; uint32_t transport_mtu_overhead; /* * transport init must allocate the new transport * and perform all internal initializations * (threads, lists, etc). */ int (*transport_init)(knet_handle_t knet_h); /* * transport free must releases _all_ resources * allocated by tranport_init */ int (*transport_free)(knet_handle_t knet_h); /* * link operations should take care of all the * sockets and epoll management for a given link/transport set * transport_link_disable should return err = -1 and errno = EBUSY * if listener is still in use, and any other errno in case * the link cannot be disabled. * * set_config/clear_config are invoked in global write lock context */ int (*transport_link_set_config)(knet_handle_t knet_h, struct knet_link *link); int (*transport_link_clear_config)(knet_handle_t knet_h, struct knet_link *link); /* * transport callback for incoming dynamic connections * this is called in global read lock context */ int (*transport_link_dyn_connect)(knet_handle_t knet_h, int sockfd, struct knet_link *link); /* * per transport error handling of recvmmsg * (see _handle_recv_from_links comments for details) */ /* * transport_rx_sock_error is invoked when recvmmsg returns <= 0 * * transport_rx_sock_error is invoked with both global_rdlock */ int (*transport_rx_sock_error)(knet_handle_t knet_h, int sockfd, int recv_err, int recv_errno); /* * transport_tx_sock_error is invoked with global_rwlock and * it's invoked when sendto or sendmmsg returns =< 0 * * it should return: * -1 on internal error * 0 ignore error and continue * 1 retry * any sleep or wait action should happen inside the transport code */ int (*transport_tx_sock_error)(knet_handle_t knet_h, int sockfd, int recv_err, int recv_errno); /* * this function is called on _every_ received packet * to verify if the packet is data or internal protocol error handling * * it should return: * -1 on error * 0 packet is not data and we should continue the packet process loop * 1 packet is not data and we should STOP the packet process loop * 2 packet is data and should be parsed as such * * transport_rx_is_data is invoked with both global_rwlock * and fd_tracker read lock (from RX thread) */ int (*transport_rx_is_data)(knet_handle_t knet_h, int sockfd, struct knet_mmsghdr *msg); } knet_transport_ops_t; socklen_t sockaddr_len(const struct sockaddr_storage *ss); /** * This is a kernel style list implementation. * * @author Steven Dake */ struct knet_list_head { struct knet_list_head *next; struct knet_list_head *prev; }; /** * @def KNET_LIST_DECLARE() * Declare and initialize a list head. */ #define KNET_LIST_DECLARE(name) \ struct knet_list_head name = { &(name), &(name) } #define KNET_INIT_LIST_HEAD(ptr) do { \ (ptr)->next = (ptr); (ptr)->prev = (ptr); \ } while (0) /** * Initialize the list entry. * * Points next and prev pointers to head. * @param head pointer to the list head */ static inline void knet_list_init(struct knet_list_head *head) { head->next = head; head->prev = head; } /** * Add this element to the list. * * @param element the new element to insert. * @param head pointer to the list head */ static inline void knet_list_add(struct knet_list_head *element, struct knet_list_head *head) { head->next->prev = element; element->next = head->next; element->prev = head; head->next = element; } /** * Add to the list (but at the end of the list). * * @param element pointer to the element to add * @param head pointer to the list head * @see knet_list_add() */ static inline void knet_list_add_tail(struct knet_list_head *element, struct knet_list_head *head) { head->prev->next = element; element->next = head; element->prev = head->prev; head->prev = element; } /** * Delete an entry from the list. * * @param _remove the list item to remove */ static inline void knet_list_del(struct knet_list_head *_remove) { _remove->next->prev = _remove->prev; _remove->prev->next = _remove->next; } /** * Replace old entry by new one * @param old: the element to be replaced * @param new: the new element to insert */ static inline void knet_list_replace(struct knet_list_head *old, struct knet_list_head *new) { new->next = old->next; new->next->prev = new; new->prev = old->prev; new->prev->next = new; } /** * Tests whether list is the last entry in list head * @param list: the entry to test * @param head: the head of the list * @return boolean true/false */ static inline int knet_list_is_last(const struct knet_list_head *list, const struct knet_list_head *head) { return list->next == head; } /** * A quick test to see if the list is empty (pointing to it's self). * @param head pointer to the list head * @return boolean true/false */ static inline int32_t knet_list_empty(const struct knet_list_head *head) { return head->next == head; } /** * Get the struct for this entry * @param ptr: the &struct list_head pointer. * @param type: the type of the struct this is embedded in. * @param member: the name of the list_struct within the struct. */ #define knet_list_entry(ptr,type,member)\ ((type *)((char *)(ptr)-(char*)(&((type *)0)->member))) /** * Get the first element from a list * @param ptr: the &struct list_head pointer. * @param type: the type of the struct this is embedded in. * @param member: the name of the list_struct within the struct. */ #define knet_list_first_entry(ptr, type, member) \ knet_list_entry((ptr)->next, type, member) /** * Iterate over a list * @param pos: the &struct list_head to use as a loop counter. * @param head: the head for your list. */ #define knet_list_for_each(pos, head) \ for (pos = (head)->next; pos != (head); pos = pos->next) /** * Iterate over a list backwards * @param pos: the &struct list_head to use as a loop counter. * @param head: the head for your list. */ #define knet_list_for_each_reverse(pos, head) \ for (pos = (head)->prev; pos != (head); pos = pos->prev) /** * Iterate over a list safe against removal of list entry * @param pos: the &struct list_head to use as a loop counter. * @param n: another &struct list_head to use as temporary storage * @param head: the head for your list. */ #define knet_list_for_each_safe(pos, n, head) \ for (pos = (head)->next, n = pos->next; pos != (head); \ pos = n, n = pos->next) /** * Iterate over list of given type * @param pos: the type * to use as a loop counter. * @param head: the head for your list. * @param member: the name of the list_struct within the struct. */ #define knet_list_for_each_entry(pos, head, member) \ for (pos = knet_list_entry((head)->next, typeof(*pos), member); \ &pos->member != (head); \ pos = knet_list_entry(pos->member.next, typeof(*pos), member)) #endif diff --git a/libknet/threads_pmtud.c b/libknet/threads_pmtud.c index 701e0fd6..4e727e91 100644 --- a/libknet/threads_pmtud.c +++ b/libknet/threads_pmtud.c @@ -1,515 +1,531 @@ /* * Copyright (C) 2015-2017 Red Hat, Inc. All rights reserved. * * Authors: Fabio M. Di Nitto * Federico Simoncelli * * This software licensed under GPL-2.0+, LGPL-2.0+ */ #include "config.h" #include #include #include #include #include "crypto.h" #include "links.h" #include "host.h" #include "logging.h" #include "transports.h" #include "threads_common.h" #include "threads_pmtud.h" static int _handle_check_link_pmtud(knet_handle_t knet_h, struct knet_host *dst_host, struct knet_link *dst_link) { int err, ret, savederrno, mutex_retry_limit, failsafe, has_kernel_mtu; uint32_t kernel_mtu; /* record kernel_mtu from EMSGSIZE */ size_t onwire_len; /* current packet onwire size */ size_t overhead_len; /* onwire packet overhead (protocol based) */ size_t max_mtu_len; /* max mtu for protocol */ size_t data_len; /* how much data we can send in the packet * generally would be onwire_len - overhead_len * needs to be adjusted for crypto */ size_t pad_len; /* crypto packet pad size, needs to move into crypto.c callbacks */ ssize_t len; /* len of what we were able to sendto onwire */ struct timespec ts; unsigned long long pong_timeout_adj_tmp; unsigned char *outbuf = (unsigned char *)knet_h->pmtudbuf; mutex_retry_limit = 0; failsafe = 0; pad_len = 0; dst_link->last_bad_mtu = 0; knet_h->pmtudbuf->khp_pmtud_link = dst_link->link_id; switch (dst_link->dst_addr.ss_family) { case AF_INET6: max_mtu_len = KNET_PMTUD_SIZE_V6; overhead_len = KNET_PMTUD_OVERHEAD_V6 + dst_link->proto_overhead; dst_link->last_good_mtu = dst_link->last_ping_size + overhead_len; break; case AF_INET: max_mtu_len = KNET_PMTUD_SIZE_V4; overhead_len = KNET_PMTUD_OVERHEAD_V4 + dst_link->proto_overhead; dst_link->last_good_mtu = dst_link->last_ping_size + overhead_len; break; default: log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD aborted, unknown protocol"); return -1; break; } /* * discovery starts from the top because kernel will * refuse to send packets > current iface mtu. * this saves us some time and network bw. */ onwire_len = max_mtu_len; restart: /* * prevent a race when interface mtu is changed _exactly_ during * the discovery process and it's complex to detect. Easier * to wait the next loop. * 30 is not an arbitrary value. To bisect from 576 to 128000 doesn't * take more than 18/19 steps. */ if (failsafe == 30) { log_err(knet_h, KNET_SUB_PMTUD, "Aborting PMTUD process: Too many attempts. MTU might have changed during discovery."); return -1; } else { failsafe++; } data_len = onwire_len - overhead_len; if (knet_h->crypto_instance) { if (knet_h->sec_block_size) { pad_len = knet_h->sec_block_size - (data_len % knet_h->sec_block_size); if (pad_len == knet_h->sec_block_size) { pad_len = 0; } data_len = data_len + pad_len; } data_len = data_len + (knet_h->sec_hash_size + knet_h->sec_salt_size + knet_h->sec_block_size); if (knet_h->sec_block_size) { while (data_len + overhead_len >= max_mtu_len) { data_len = data_len - knet_h->sec_block_size; } } if (dst_link->last_bad_mtu) { while (data_len + overhead_len >= dst_link->last_bad_mtu) { data_len = data_len - (knet_h->sec_hash_size + knet_h->sec_salt_size + knet_h->sec_block_size); } } if (data_len < (knet_h->sec_hash_size + knet_h->sec_salt_size + knet_h->sec_block_size) + 1) { log_debug(knet_h, KNET_SUB_PMTUD, "Aborting PMTUD process: link mtu smaller than crypto header detected (link might have been disconnected)"); return -1; } onwire_len = data_len + overhead_len; knet_h->pmtudbuf->khp_pmtud_size = onwire_len; if (crypto_encrypt_and_sign(knet_h, (const unsigned char *)knet_h->pmtudbuf, data_len - (knet_h->sec_hash_size + knet_h->sec_salt_size + knet_h->sec_block_size), knet_h->pmtudbuf_crypt, (ssize_t *)&data_len) < 0) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to crypto pmtud packet"); return -1; } outbuf = knet_h->pmtudbuf_crypt; knet_h->stats_extra.tx_crypt_pmtu_packets++; } else { knet_h->pmtudbuf->khp_pmtud_size = onwire_len; } /* link has gone down, aborting pmtud */ if (dst_link->status.connected != 1) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD detected host (%u) link (%u) has been disconnected", dst_host->host_id, dst_link->link_id); return -1; } if (dst_link->transport_connected != 1) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD detected host (%u) link (%u) has been disconnected", dst_host->host_id, dst_link->link_id); return -1; } if (pthread_mutex_lock(&knet_h->pmtud_mutex) != 0) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get mutex lock"); return -1; } if (knet_h->pmtud_abort) { pthread_mutex_unlock(&knet_h->pmtud_mutex); errno = EDEADLK; return -1; } savederrno = pthread_mutex_lock(&knet_h->tx_mutex); if (savederrno) { log_err(knet_h, KNET_SUB_PMTUD, "Unable to get TX mutex lock: %s", strerror(savederrno)); return -1; } retry: len = sendto(dst_link->outsock, outbuf, data_len, MSG_DONTWAIT | MSG_NOSIGNAL, (struct sockaddr *) &dst_link->dst_addr, sizeof(struct sockaddr_storage)); savederrno = errno; if (pthread_mutex_lock(&knet_h->kmtu_mutex) != 0) { has_kernel_mtu = 0; } else { has_kernel_mtu = 1; knet_h->kernel_mtu = 0; pthread_mutex_unlock(&knet_h->kmtu_mutex); } kernel_mtu = 0; err = transport_tx_sock_error(knet_h, dst_link->transport_type, dst_link->outsock, len, savederrno); switch(err) { case -1: /* unrecoverable error */ log_debug(knet_h, KNET_SUB_PMTUD, "Unable to send pmtu packet (sendto): %d %s", savederrno, strerror(savederrno)); pthread_mutex_unlock(&knet_h->tx_mutex); pthread_mutex_unlock(&knet_h->pmtud_mutex); dst_link->status.stats.tx_pmtu_errors++; return -1; case 0: /* ignore error and continue */ break; case 1: /* retry to send those same data */ dst_link->status.stats.tx_pmtu_retries++; goto retry; break; } pthread_mutex_unlock(&knet_h->tx_mutex); if (len != (ssize_t )data_len) { if (savederrno == EMSGSIZE) { if (has_kernel_mtu) { if (pthread_mutex_lock(&knet_h->kmtu_mutex) != 0) { has_kernel_mtu = 0; } else { kernel_mtu = knet_h->kernel_mtu; pthread_mutex_unlock(&knet_h->kmtu_mutex); } } if (kernel_mtu > 0) { dst_link->last_bad_mtu = kernel_mtu + 1; } else { dst_link->last_bad_mtu = onwire_len; } } else { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to send pmtu packet len: %zu err: %s", onwire_len, strerror(savederrno)); } } else { dst_link->last_sent_mtu = onwire_len; dst_link->last_recv_mtu = 0; dst_link->status.stats.tx_pmtu_packets++; dst_link->status.stats.tx_pmtu_bytes += data_len; if (clock_gettime(CLOCK_REALTIME, &ts) < 0) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get current time: %s", strerror(errno)); pthread_mutex_unlock(&knet_h->pmtud_mutex); return -1; } /* * set PMTUd reply timeout to match pong_timeout on a given link * * math: internally pong_timeout is expressed in microseconds, while * the public API exports milliseconds. So careful with the 0's here. * the loop is necessary because we are grabbing the current time just above * and add values to it that could overflow into seconds. */ if (pthread_mutex_lock(&knet_h->backoff_mutex)) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get backoff_mutex"); pthread_mutex_unlock(&knet_h->pmtud_mutex); return -1; } if (knet_h->crypto_instance) { /* * crypto, under pressure, is a royal PITA */ pong_timeout_adj_tmp = dst_link->pong_timeout_adj * 2; } else { pong_timeout_adj_tmp = dst_link->pong_timeout_adj; } ts.tv_sec += pong_timeout_adj_tmp / 1000000; ts.tv_nsec += (((pong_timeout_adj_tmp) % 1000000) * 1000); while (ts.tv_nsec > 1000000000) { ts.tv_sec += 1; ts.tv_nsec -= 1000000000; } pthread_mutex_unlock(&knet_h->backoff_mutex); knet_h->pmtud_waiting = 1; ret = pthread_cond_timedwait(&knet_h->pmtud_cond, &knet_h->pmtud_mutex, &ts); knet_h->pmtud_waiting = 0; if (knet_h->pmtud_abort) { pthread_mutex_unlock(&knet_h->pmtud_mutex); errno = EDEADLK; return -1; } if (shutdown_in_progress(knet_h)) { pthread_mutex_unlock(&knet_h->pmtud_mutex); log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD aborted. shutdown in progress"); return -1; } if ((ret != 0) && (ret != ETIMEDOUT)) { pthread_mutex_unlock(&knet_h->pmtud_mutex); if (mutex_retry_limit == 3) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD aborted, unable to get mutex lock"); return -1; } mutex_retry_limit++; goto restart; } if ((dst_link->last_recv_mtu != onwire_len) || (ret)) { dst_link->last_bad_mtu = onwire_len; } else { int found_mtu = 0; if (knet_h->sec_block_size) { if ((onwire_len + knet_h->sec_block_size >= max_mtu_len) || ((dst_link->last_bad_mtu) && (dst_link->last_bad_mtu <= (onwire_len + knet_h->sec_block_size)))) { found_mtu = 1; } } else { if ((onwire_len == max_mtu_len) || ((dst_link->last_bad_mtu) && (dst_link->last_bad_mtu == (onwire_len + 1))) || (dst_link->last_bad_mtu == dst_link->last_good_mtu)) { found_mtu = 1; } } if (found_mtu) { /* * account for IP overhead, knet headers and crypto in PMTU calculation */ dst_link->status.mtu = onwire_len - dst_link->status.proto_overhead; pthread_mutex_unlock(&knet_h->pmtud_mutex); return 0; } dst_link->last_good_mtu = onwire_len; } } if (kernel_mtu) { onwire_len = kernel_mtu; } else { onwire_len = (dst_link->last_good_mtu + dst_link->last_bad_mtu) / 2; } pthread_mutex_unlock(&knet_h->pmtud_mutex); /* * give time to the kernel to determine its own version of MTU */ usleep((dst_link->status.stats.latency_ave * 1000) / 4); goto restart; } -static int _handle_check_pmtud(knet_handle_t knet_h, struct knet_host *dst_host, struct knet_link *dst_link, unsigned int *min_mtu) +static int _handle_check_pmtud(knet_handle_t knet_h, struct knet_host *dst_host, struct knet_link *dst_link, unsigned int *min_mtu, int force_run) { uint8_t saved_valid_pmtud; unsigned int saved_pmtud; struct timespec clock_now; unsigned long long diff_pmtud, interval; - interval = knet_h->pmtud_interval * 1000000000llu; /* nanoseconds */ + if (!force_run) { + interval = knet_h->pmtud_interval * 1000000000llu; /* nanoseconds */ - if (clock_gettime(CLOCK_MONOTONIC, &clock_now) != 0) { - log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get monotonic clock"); - return 0; - } + if (clock_gettime(CLOCK_MONOTONIC, &clock_now) != 0) { + log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get monotonic clock"); + return 0; + } - timespec_diff(dst_link->pmtud_last, clock_now, &diff_pmtud); + timespec_diff(dst_link->pmtud_last, clock_now, &diff_pmtud); - if (diff_pmtud < interval) { - *min_mtu = dst_link->status.mtu; - return dst_link->has_valid_mtu; + if (diff_pmtud < interval) { + *min_mtu = dst_link->status.mtu; + return dst_link->has_valid_mtu; + } } switch (dst_link->dst_addr.ss_family) { case AF_INET6: dst_link->status.proto_overhead = KNET_PMTUD_OVERHEAD_V6 + dst_link->proto_overhead + KNET_HEADER_ALL_SIZE + knet_h->sec_header_size; break; case AF_INET: dst_link->status.proto_overhead = KNET_PMTUD_OVERHEAD_V4 + dst_link->proto_overhead + KNET_HEADER_ALL_SIZE + knet_h->sec_header_size; break; } saved_pmtud = dst_link->status.mtu; saved_valid_pmtud = dst_link->has_valid_mtu; log_debug(knet_h, KNET_SUB_PMTUD, "Starting PMTUD for host: %u link: %u", dst_host->host_id, dst_link->link_id); if (_handle_check_link_pmtud(knet_h, dst_host, dst_link) < 0) { if (errno == EDEADLK) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD for host: %u link: %u has been rescheduled", dst_host->host_id, dst_link->link_id); dst_link->status.mtu = saved_pmtud; dst_link->has_valid_mtu = saved_valid_pmtud; errno = EDEADLK; return dst_link->has_valid_mtu; } dst_link->has_valid_mtu = 0; } else { dst_link->has_valid_mtu = 1; switch (dst_link->dst_addr.ss_family) { case AF_INET6: if (((dst_link->status.mtu + dst_link->status.proto_overhead) < KNET_PMTUD_MIN_MTU_V6) || ((dst_link->status.mtu + dst_link->status.proto_overhead) > KNET_PMTUD_SIZE_V6)) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD detected an IPv6 MTU out of bound value (%u) for host: %u link: %u.", dst_link->status.mtu + dst_link->status.proto_overhead, dst_host->host_id, dst_link->link_id); dst_link->has_valid_mtu = 0; } break; case AF_INET: if (((dst_link->status.mtu + dst_link->status.proto_overhead) < KNET_PMTUD_MIN_MTU_V4) || ((dst_link->status.mtu + dst_link->status.proto_overhead) > KNET_PMTUD_SIZE_V4)) { log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD detected an IPv4 MTU out of bound value (%u) for host: %u link: %u.", dst_link->status.mtu + dst_link->status.proto_overhead, dst_host->host_id, dst_link->link_id); dst_link->has_valid_mtu = 0; } break; } if (dst_link->has_valid_mtu) { if ((saved_pmtud) && (saved_pmtud != dst_link->status.mtu)) { log_info(knet_h, KNET_SUB_PMTUD, "PMTUD link change for host: %u link: %u from %u to %u", dst_host->host_id, dst_link->link_id, saved_pmtud, dst_link->status.mtu); } log_debug(knet_h, KNET_SUB_PMTUD, "PMTUD completed for host: %u link: %u current link mtu: %u", dst_host->host_id, dst_link->link_id, dst_link->status.mtu); if (dst_link->status.mtu < *min_mtu) { *min_mtu = dst_link->status.mtu; } /* * set pmtud_last, if we can, after we are done with the PMTUd process * because it can take a very long time. */ dst_link->pmtud_last = clock_now; if (!clock_gettime(CLOCK_MONOTONIC, &clock_now)) { dst_link->pmtud_last = clock_now; } } } if (saved_valid_pmtud != dst_link->has_valid_mtu) { _host_dstcache_update_sync(knet_h, dst_host); } return dst_link->has_valid_mtu; } void *_handle_pmtud_link_thread(void *data) { knet_handle_t knet_h = (knet_handle_t) data; struct knet_host *dst_host; struct knet_link *dst_link; int link_idx; unsigned int min_mtu, have_mtu; unsigned int lower_mtu; int link_has_mtu; + int force_run = 0; knet_h->data_mtu = KNET_PMTUD_MIN_MTU_V4 - KNET_HEADER_ALL_SIZE - knet_h->sec_header_size; /* preparing pmtu buffer */ knet_h->pmtudbuf->kh_version = KNET_HEADER_VERSION; knet_h->pmtudbuf->kh_type = KNET_HEADER_TYPE_PMTUD; knet_h->pmtudbuf->kh_node = htons(knet_h->host_id); while (!shutdown_in_progress(knet_h)) { usleep(KNET_THREADS_TIMERES); if (pthread_mutex_lock(&knet_h->pmtud_mutex) != 0) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get mutex lock"); continue; } knet_h->pmtud_abort = 0; + knet_h->pmtud_running = 1; + force_run = knet_h->pmtud_forcerun; + knet_h->pmtud_forcerun = 0; pthread_mutex_unlock(&knet_h->pmtud_mutex); + if (force_run) { + log_debug(knet_h, KNET_SUB_PMTUD, "PMTUd request to rerun has been received"); + } + if (pthread_rwlock_rdlock(&knet_h->global_rwlock) != 0) { log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get read lock"); continue; } lower_mtu = KNET_PMTUD_SIZE_V4; min_mtu = KNET_PMTUD_SIZE_V4 - KNET_HEADER_ALL_SIZE - knet_h->sec_header_size; have_mtu = 0; for (dst_host = knet_h->host_head; dst_host != NULL; dst_host = dst_host->next) { for (link_idx = 0; link_idx < KNET_MAX_LINK; link_idx++) { dst_link = &dst_host->link[link_idx]; if ((dst_link->status.enabled != 1) || (dst_link->status.connected != 1) || (dst_host->link[link_idx].transport_type == KNET_TRANSPORT_LOOPBACK) || (!dst_link->last_ping_size) || ((dst_link->dynamic == KNET_LINK_DYNIP) && (dst_link->status.dynconnected != 1))) continue; - link_has_mtu = _handle_check_pmtud(knet_h, dst_host, dst_link, &min_mtu); + link_has_mtu = _handle_check_pmtud(knet_h, dst_host, dst_link, &min_mtu, force_run); if (errno == EDEADLK) { goto out_unlock; } if (link_has_mtu) { have_mtu = 1; if (min_mtu < lower_mtu) { lower_mtu = min_mtu; } } } } if (have_mtu) { if (knet_h->data_mtu != lower_mtu) { knet_h->data_mtu = lower_mtu; log_info(knet_h, KNET_SUB_PMTUD, "Global data MTU changed to: %u", knet_h->data_mtu); if (knet_h->pmtud_notify_fn) { knet_h->pmtud_notify_fn(knet_h->pmtud_notify_fn_private_data, knet_h->data_mtu); } } } out_unlock: pthread_rwlock_unlock(&knet_h->global_rwlock); + if (pthread_mutex_lock(&knet_h->pmtud_mutex) != 0) { + log_debug(knet_h, KNET_SUB_PMTUD, "Unable to get mutex lock"); + } else { + knet_h->pmtud_running = 0; + pthread_mutex_unlock(&knet_h->pmtud_mutex); + } } return NULL; } diff --git a/libknet/transport_udp.c b/libknet/transport_udp.c index 4b2826b2..aec42d07 100644 --- a/libknet/transport_udp.c +++ b/libknet/transport_udp.c @@ -1,411 +1,428 @@ /* * Copyright (C) 2016-2017 Red Hat, Inc. All rights reserved. * * Author: Christine Caulfield * * This software licensed under GPL-2.0+, LGPL-2.0+ */ #include "config.h" #include #include #include #include #include #include #include #include #include #if defined (IP_RECVERR) || defined (IPV6_RECVERR) #include #endif #include "libknet.h" #include "compat.h" #include "host.h" #include "link.h" #include "logging.h" #include "common.h" #include "transport_common.h" #include "transport_udp.h" #include "threads_common.h" typedef struct udp_handle_info { struct knet_list_head links_list; } udp_handle_info_t; typedef struct udp_link_info { struct knet_list_head list; struct sockaddr_storage local_address; int socket_fd; int on_epoll; } udp_link_info_t; int udp_transport_link_set_config(knet_handle_t knet_h, struct knet_link *kn_link) { int err = 0, savederrno = 0; int sock = -1; struct epoll_event ev; udp_link_info_t *info; udp_handle_info_t *handle_info = knet_h->transports[KNET_TRANSPORT_UDP]; #if defined (IP_RECVERR) || defined (IPV6_RECVERR) int value; #endif /* * Only allocate a new link if the local address is different */ knet_list_for_each_entry(info, &handle_info->links_list, list) { if (memcmp(&info->local_address, &kn_link->src_addr, sizeof(struct sockaddr_storage)) == 0) { log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Re-using existing UDP socket for new link"); kn_link->outsock = info->socket_fd; kn_link->transport_link = info; kn_link->transport_connected = 1; return 0; } } info = malloc(sizeof(udp_link_info_t)); if (!info) { err = -1; goto exit_error; } sock = socket(kn_link->src_addr.ss_family, SOCK_DGRAM, 0); if (sock < 0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to create listener socket: %s", strerror(savederrno)); goto exit_error; } if (_configure_transport_socket(knet_h, sock, &kn_link->src_addr, kn_link->flags, "UDP") < 0) { savederrno = errno; err = -1; goto exit_error; } #ifdef IP_RECVERR if (kn_link->src_addr.ss_family == AF_INET) { value = 1; if (setsockopt(sock, SOL_IP, IP_RECVERR, &value, sizeof(value)) <0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to set RECVERR on socket: %s", strerror(savederrno)); goto exit_error; } log_debug(knet_h, KNET_SUB_TRANSP_UDP, "IP_RECVERR enabled on socket: %i", sock); } #else log_debug(knet_h, KNET_SUB_TRANSP_UDP, "IP_RECVERR not available in this build/platform"); #endif #ifdef IPV6_RECVERR if (kn_link->src_addr.ss_family == AF_INET6) { value = 1; if (setsockopt(sock, SOL_IPV6, IPV6_RECVERR, &value, sizeof(value)) <0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to set RECVERR on socket: %s", strerror(savederrno)); goto exit_error; } log_debug(knet_h, KNET_SUB_TRANSP_UDP, "IPV6_RECVERR enabled on socket: %i", sock); } #else log_debug(knet_h, KNET_SUB_TRANSP_UDP, "IPV6_RECVERR not available in this build/platform"); #endif if (bind(sock, (struct sockaddr *)&kn_link->src_addr, sockaddr_len(&kn_link->src_addr))) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to bind listener socket: %s", strerror(savederrno)); goto exit_error; } memset(&ev, 0, sizeof(struct epoll_event)); ev.events = EPOLLIN; ev.data.fd = sock; if (epoll_ctl(knet_h->recv_from_links_epollfd, EPOLL_CTL_ADD, sock, &ev)) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to add listener to epoll pool: %s", strerror(savederrno)); goto exit_error; } info->on_epoll = 1; if (_set_fd_tracker(knet_h, sock, KNET_TRANSPORT_UDP, 0, info) < 0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to set fd tracker: %s", strerror(savederrno)); goto exit_error; } memmove(&info->local_address, &kn_link->src_addr, sizeof(struct sockaddr_storage)); info->socket_fd = sock; knet_list_add(&info->list, &handle_info->links_list); kn_link->outsock = sock; kn_link->transport_link = info; kn_link->transport_connected = 1; exit_error: if (err) { if (info) { if (info->on_epoll) { epoll_ctl(knet_h->recv_from_links_epollfd, EPOLL_CTL_DEL, sock, &ev); } free(info); } if (sock >= 0) { close(sock); } } errno = savederrno; return err; } int udp_transport_link_clear_config(knet_handle_t knet_h, struct knet_link *kn_link) { int err = 0, savederrno = 0; int found = 0; struct knet_host *host; int link_idx; udp_link_info_t *info = kn_link->transport_link; struct epoll_event ev; for (host = knet_h->host_head; host != NULL; host = host->next) { for (link_idx = 0; link_idx < KNET_MAX_LINK; link_idx++) { if (&host->link[link_idx] == kn_link) continue; if ((host->link[link_idx].transport_link == info) && (host->link[link_idx].status.enabled == 1)) { found = 1; break; } } } if (found) { log_debug(knet_h, KNET_SUB_TRANSP_UDP, "UDP socket %d still in use", info->socket_fd); savederrno = EBUSY; err = -1; goto exit_error; } if (info->on_epoll) { memset(&ev, 0, sizeof(struct epoll_event)); ev.events = EPOLLIN; ev.data.fd = info->socket_fd; if (epoll_ctl(knet_h->recv_from_links_epollfd, EPOLL_CTL_DEL, info->socket_fd, &ev) < 0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to remove UDP socket from epoll poll: %s", strerror(errno)); goto exit_error; } info->on_epoll = 0; } if (_set_fd_tracker(knet_h, info->socket_fd, KNET_MAX_TRANSPORTS, 0, NULL) < 0) { savederrno = errno; err = -1; log_err(knet_h, KNET_SUB_TRANSP_UDP, "Unable to set fd tracker: %s", strerror(savederrno)); goto exit_error; } close(info->socket_fd); knet_list_del(&info->list); free(kn_link->transport_link); exit_error: errno = savederrno; return err; } int udp_transport_free(knet_handle_t knet_h) { udp_handle_info_t *handle_info; if (!knet_h->transports[KNET_TRANSPORT_UDP]) { errno = EINVAL; return -1; } handle_info = knet_h->transports[KNET_TRANSPORT_UDP]; /* * keep it here while we debug list usage and such */ if (!knet_list_empty(&handle_info->links_list)) { log_err(knet_h, KNET_SUB_TRANSP_UDP, "Internal error. handle list is not empty"); return -1; } free(handle_info); knet_h->transports[KNET_TRANSPORT_UDP] = NULL; return 0; } int udp_transport_init(knet_handle_t knet_h) { udp_handle_info_t *handle_info; if (knet_h->transports[KNET_TRANSPORT_UDP]) { errno = EEXIST; return -1; } handle_info = malloc(sizeof(udp_handle_info_t)); if (!handle_info) { return -1; } memset(handle_info, 0, sizeof(udp_handle_info_t)); knet_h->transports[KNET_TRANSPORT_UDP] = handle_info; knet_list_init(&handle_info->links_list); return 0; } #if defined (IP_RECVERR) || defined (IPV6_RECVERR) static int read_errs_from_sock(knet_handle_t knet_h, int sockfd) { int err = 0, savederrno = 0; int got_err = 0; char buffer[1024]; struct iovec iov; struct msghdr msg; struct cmsghdr *cmsg; struct sock_extended_err *sock_err; struct icmphdr icmph; struct sockaddr_storage remote; struct sockaddr_storage *origin; char addr_str[KNET_MAX_HOST_LEN]; char port_str[KNET_MAX_PORT_LEN]; iov.iov_base = &icmph; iov.iov_len = sizeof(icmph); msg.msg_name = (void*)&remote; msg.msg_namelen = sizeof(remote); msg.msg_iov = &iov; msg.msg_iovlen = 1; msg.msg_flags = 0; msg.msg_control = buffer; msg.msg_controllen = sizeof(buffer); for (;;) { err = recvmsg(sockfd, &msg, MSG_ERRQUEUE); savederrno = errno; if (err < 0) { if (!got_err) { errno = savederrno; return -1; } else { return 0; } } got_err = 1; for (cmsg = CMSG_FIRSTHDR(&msg);cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) { if (((cmsg->cmsg_level == SOL_IP) && (cmsg->cmsg_type == IP_RECVERR)) || ((cmsg->cmsg_level == SOL_IPV6 && (cmsg->cmsg_type == IPV6_RECVERR)))) { sock_err = (struct sock_extended_err*)(void *)CMSG_DATA(cmsg); if (sock_err) { switch (sock_err->ee_origin) { case 0: /* no origin */ case 1: /* local source (EMSGSIZE) */ if (sock_err->ee_errno == EMSGSIZE) { if (pthread_mutex_lock(&knet_h->kmtu_mutex) != 0) { + log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Unable to get mutex lock"); knet_h->kernel_mtu = 0; + break; } else { knet_h->kernel_mtu = sock_err->ee_info; pthread_mutex_unlock(&knet_h->kmtu_mutex); } + + /* + * we can only try to take a lock here. This part of the code + * can be invoked by any thread, including PMTUd that is already + * holding a lock at that stage. + * If PMTUd is holding the lock, mostlikely it is already running + * and we don't need to notify it back. + */ + if (!pthread_mutex_trylock(&knet_h->pmtud_mutex)) { + if (!knet_h->pmtud_running) { + log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Notifying PMTUd to rerun"); + knet_h->pmtud_forcerun = 1; + } + pthread_mutex_unlock(&knet_h->pmtud_mutex); + } } /* * those errors are way too noisy */ break; case 2: /* ICMP */ case 3: /* ICMP6 */ origin = (struct sockaddr_storage *)(void *)SO_EE_OFFENDER(sock_err); if (knet_addrtostr(origin, sizeof(origin), addr_str, KNET_MAX_HOST_LEN, port_str, KNET_MAX_PORT_LEN) < 0) { log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Received ICMP error from unknown source: %s", strerror(sock_err->ee_errno)); } else { log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Received ICMP error from %s: %s", addr_str, strerror(sock_err->ee_errno)); } break; } } else { log_debug(knet_h, KNET_SUB_TRANSP_UDP, "No data in MSG_ERRQUEUE"); } } } } } #else static int read_errs_from_sock(knet_handle_t knet_h, int sockfd) { return 0; } #endif int udp_transport_rx_sock_error(knet_handle_t knet_h, int sockfd, int recv_err, int recv_errno) { if (recv_errno == EAGAIN) { read_errs_from_sock(knet_h, sockfd); } return 0; } int udp_transport_tx_sock_error(knet_handle_t knet_h, int sockfd, int recv_err, int recv_errno) { if (recv_err < 0) { if (recv_errno == EMSGSIZE) { read_errs_from_sock(knet_h, sockfd); return 0; } if ((recv_errno == ENOBUFS) || (recv_errno == EAGAIN)) { #ifdef DEBUG log_debug(knet_h, KNET_SUB_TRANSP_UDP, "Sock: %d is overloaded. Slowing TX down", sockfd); #endif usleep(KNET_THREADS_TIMERES / 16); } else { read_errs_from_sock(knet_h, sockfd); } return 1; } return 0; } int udp_transport_rx_is_data(knet_handle_t knet_h, int sockfd, struct knet_mmsghdr *msg) { if (msg->msg_len == 0) return 0; return 2; } int udp_transport_link_dyn_connect(knet_handle_t knet_h, int sockfd, struct knet_link *kn_link) { kn_link->status.dynconnected = 1; return 0; }