Re: [PATCH v3 4/6] tcp, flow: Replace per-connection in_epoll flag with threadnb in flow_common

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 16599 bytes --]

On Thu, Oct 09, 2025 at 03:04:07PM +0200, Laurent Vivier wrote:
> The in_epoll boolean flag in tcp_tap_conn and tcp_splice_conn only tracked
> whether a connection was registered with epoll, not which epoll instance.
> This limited flexibility for future multi-epoll support.
> 
> Replace the boolean with a threadnb field in flow_common that identifies
> which thread (and thus which epoll instance) the flow is registered with.
> Use FLOW_THREADNB_INVALID to indicate when a flow is not registered with
> any epoll instance. A threadnb_to_epollfd[] mapping table translates
> thread numbers to their corresponding epoll file descriptors.
> 
> Add helper functions:
> - flow_epollfd_valid() to check if a flow is registered with epoll
> - flow_epollfd_get() to retrieve the epoll fd for a flow's thread
> - flow_epollfd_set() to register a flow with a thread's epoll instance
> 
> This change also simplifies tcp_timer_ctl() and conn_flag_do() by removing
> the need to pass the context 'c', since the epoll fd is now directly
> accessible from the flow structure via flow_epollfd_get().
> 
> Signed-off-by: Laurent Vivier <lvivier@redhat.com>

I have mixed feelings on the naming.  "threadnb" suggests the intended
use for this for multithreading.  However, this mechanism doesn't
inherently have to do with threading.  Not sure which consideration is
more important here.

> ---
>  flow.c       | 43 ++++++++++++++++++++++++++++++++++++++++++-
>  flow.h       | 11 +++++++++++
>  tcp.c        | 39 +++++++++++++++++++++------------------
>  tcp_conn.h   |  8 +-------
>  tcp_splice.c | 24 ++++++++++++------------
>  5 files changed, 87 insertions(+), 38 deletions(-)
> 
> diff --git a/flow.c b/flow.c
> index b14e9d8b63ff..6298bbafcd51 100644
> --- a/flow.c
> +++ b/flow.c
> @@ -116,6 +116,7 @@ static_assert(ARRAY_SIZE(flow_proto) == FLOW_NUM_TYPES,
>  unsigned flow_first_free;
>  union flow flowtab[FLOW_MAX];
>  static const union flow *flow_new_entry; /* = NULL */
> +static int threadnb_to_epollfd[FLOW_THREADNB_SIZE];
>  
>  /* Hash table to index it */
>  #define FLOW_HASH_LOAD		70		/* % */
> @@ -347,6 +348,45 @@ static void flow_set_state(struct flow_common *f, enum flow_state state)
>  	flow_log_details_(f, LOG_DEBUG, MAX(state, oldstate));
>  }
>  
> +/**
> + * flow_epollfd_valid() - Check if flow is registered with an epoll instance

I don't think "valid" really conveys what this does.  Maybe
flow_in_epoll(), since it reproduces the old in_epoll flag?

> + * @f:		Flow to check
> + *
> + * Return: true if flow is registered with epoll, false otherwise
> + */
> +bool flow_epollfd_valid(const struct flow_common *f)
> +{
> +	return f->threadnb != FLOW_THREADNB_INVALID;
> +}
> +
> +/**
> + * flow_epollfd_get() - Get the epoll file descriptor for a flow

flow_epollfd().  The "get" doesn't really add anything.

> + * @f:		Flow to query
> + *
> + * Return: epoll file descriptor associated with the flow's thread
> + */
> +int flow_epollfd_get(const struct flow_common *f)
> +{
> +	ASSERT(f->threadnb < FLOW_THREADNB_MAX);
> +
> +	return threadnb_to_epollfd[f->threadnb];
> +}
> +
> +/**
> + * flow_epollfd_set() - Set the epoll instance for a flow
> + * @f:		Flow to update
> + * @threadnb:	Thread number to associate with this flow
> + * @epollfd:	epoll file descriptor for the thread
> + */
> +void flow_epollfd_set(struct flow_common *f, int threadnb, int epollfd)
> +{
> +	ASSERT(threadnb < FLOW_THREADNB_MAX);
> +	ASSERT(epollfd >= 0);
> +
> +	threadnb_to_epollfd[threadnb] = epollfd;
> +	f->threadnb = threadnb;

This feels like it's mixing concerns: the threadb_to_epollfd[] table
should only be populated once, but each thread will need to have its
threadnb initialised.  It works as long as its always called with
matching threadnb and epollfd, but it seems like a footgun.

> +}
> +
>  /**
>   * flow_initiate_() - Move flow to INI, setting pif[INISIDE]
>   * @flow:	Flow to change state
> @@ -548,6 +588,7 @@ union flow *flow_alloc(void)
>  
>  	flow_new_entry = flow;
>  	memset(flow, 0, sizeof(*flow));
> +	flow->f.threadnb = FLOW_THREADNB_INVALID;
>  	flow_set_state(&flow->f, FLOW_STATE_NEW);
>  
>  	return flow;
> @@ -827,7 +868,7 @@ void flow_defer_handler(const struct ctx *c, const struct timespec *now)
>  		case FLOW_TCP_SPLICE:
>  			closed = tcp_splice_flow_defer(&flow->tcp_splice);
>  			if (!closed && timer)
> -				tcp_splice_timer(c, &flow->tcp_splice);
> +				tcp_splice_timer(&flow->tcp_splice);
>  			break;
>  		case FLOW_PING4:
>  		case FLOW_PING6:
> diff --git a/flow.h b/flow.h
> index ef138b83add8..18ae45d67720 100644
> --- a/flow.h
> +++ b/flow.h
> @@ -177,6 +177,8 @@ int flowside_connect(const struct ctx *c, int s,
>   * @type:	Type of packet flow
>   * @pif[]:	Interface for each side of the flow
>   * @side[]:	Information for each side of the flow
> + * @threadnb:	Thread number flow is registered with
> + *		(FLOW_THREADNB_INVALID if not)
>   */
>  struct flow_common {
>  #ifdef __GNUC__
> @@ -192,8 +194,14 @@ struct flow_common {
>  #endif
>  	uint8_t		pif[SIDES];
>  	struct flowside	side[SIDES];
> +#define FLOW_THREADNB_BITS 8
> +	unsigned int	threadnb:FLOW_THREADNB_BITS;
>  };
>  
> +#define FLOW_THREADNB_SIZE	(1 << FLOW_THREADNB_BITS)
> +#define FLOW_THREADNB_MAX	(FLOW_THREADNB_SIZE - 1)
> +#define FLOW_THREADNB_INVALID	FLOW_THREADNB_MAX
> +
>  #define FLOW_INDEX_BITS		17	/* 128k - 1 */
>  #define FLOW_MAX		MAX_FROM_BITS(FLOW_INDEX_BITS)
>  
> @@ -249,6 +257,9 @@ flow_sidx_t flow_lookup_sa(const struct ctx *c, uint8_t proto, uint8_t pif,
>  union flow;
>  
>  void flow_init(void);
> +bool flow_epollfd_valid(const struct flow_common *f);
> +int flow_epollfd_get(const struct flow_common *f);
> +void flow_epollfd_set(struct flow_common *f, int threadnb, int epollfd);
>  void flow_defer_handler(const struct ctx *c, const struct timespec *now);
>  int flow_migrate_source_early(struct ctx *c, const struct migrate_stage *stage,
>  			      int fd);
> diff --git a/tcp.c b/tcp.c
> index c2d842bbdf4f..f9c67e2b006d 100644
> --- a/tcp.c
> +++ b/tcp.c
> @@ -504,25 +504,27 @@ static uint32_t tcp_conn_epoll_events(uint8_t events, uint8_t conn_flags)
>   */
>  static int tcp_epoll_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
>  {
> -	int m = conn->in_epoll ? EPOLL_CTL_MOD : EPOLL_CTL_ADD;
> +	int m = flow_epollfd_valid(&conn->f) ? EPOLL_CTL_MOD : EPOLL_CTL_ADD;
>  	union epoll_ref ref = { .type = EPOLL_TYPE_TCP, .fd = conn->sock,
>  		                .flowside = FLOW_SIDX(conn, !TAPSIDE(conn)), };
>  	struct epoll_event ev = { .data.u64 = ref.u64 };
> +	int epollfd = flow_epollfd_valid(&conn->f) ? flow_epollfd_get(&conn->f)
> +						   : c->epollfd;
>  
>  	if (conn->events == CLOSED) {
> -		if (conn->in_epoll)
> -			epoll_del(c->epollfd, conn->sock);
> +		if (flow_epollfd_valid(&conn->f))
> +			epoll_del(epollfd, conn->sock);
>  		if (conn->timer != -1)
> -			epoll_del(c->epollfd, conn->timer);
> +			epoll_del(epollfd, conn->timer);
>  		return 0;
>  	}
>  
>  	ev.events = tcp_conn_epoll_events(conn->events, conn->flags);
>  
> -	if (epoll_ctl(c->epollfd, m, conn->sock, &ev))
> +	if (epoll_ctl(epollfd, m, conn->sock, &ev))
>  		return -errno;
>  
> -	conn->in_epoll = true;
> +	flow_epollfd_set(&conn->f, 0, epollfd);
>  
>  	if (conn->timer != -1) {
>  		union epoll_ref ref_t = { .type = EPOLL_TYPE_TCP_TIMER,
> @@ -531,7 +533,8 @@ static int tcp_epoll_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
>  		struct epoll_event ev_t = { .data.u64 = ref_t.u64,
>  					    .events = EPOLLIN | EPOLLET };
>  
> -		if (epoll_ctl(c->epollfd, EPOLL_CTL_MOD, conn->timer, &ev_t))
> +		if (epoll_ctl(flow_epollfd_get(&conn->f), EPOLL_CTL_MOD,
> +			      conn->timer, &ev_t))
>  			return -errno;
>  	}
>  
> @@ -540,12 +543,11 @@ static int tcp_epoll_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
>  
>  /**
>   * tcp_timer_ctl() - Set timerfd based on flags/events, create timerfd if needed
> - * @c:		Execution context
>   * @conn:	Connection pointer
>   *
>   * #syscalls timerfd_create timerfd_settime
>   */
> -static void tcp_timer_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
> +static void tcp_timer_ctl(struct tcp_tap_conn *conn)
>  {
>  	struct itimerspec it = { { 0 }, { 0 } };
>  
> @@ -570,7 +572,8 @@ static void tcp_timer_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
>  		}
>  		conn->timer = fd;
>  
> -		if (epoll_ctl(c->epollfd, EPOLL_CTL_ADD, conn->timer, &ev)) {
> +		if (epoll_ctl(flow_epollfd_get(&conn->f), EPOLL_CTL_ADD,
> +			      conn->timer, &ev)) {
>  			flow_dbg_perror(conn, "failed to add timer");
>  			close(conn->timer);
>  			conn->timer = -1;
> @@ -628,7 +631,7 @@ void conn_flag_do(const struct ctx *c, struct tcp_tap_conn *conn,
>  			 * flags and factor this into the logic below.
>  			 */
>  			if (flag == ACK_FROM_TAP_DUE)
> -				tcp_timer_ctl(c, conn);
> +				tcp_timer_ctl(conn);
>  
>  			return;
>  		}
> @@ -644,7 +647,7 @@ void conn_flag_do(const struct ctx *c, struct tcp_tap_conn *conn,
>  	if (flag == ACK_FROM_TAP_DUE || flag == ACK_TO_TAP_DUE		  ||
>  	    (flag == ~ACK_FROM_TAP_DUE && (conn->flags & ACK_TO_TAP_DUE)) ||
>  	    (flag == ~ACK_TO_TAP_DUE   && (conn->flags & ACK_FROM_TAP_DUE)))
> -		tcp_timer_ctl(c, conn);
> +		tcp_timer_ctl(conn);
>  }
>  
>  /**
> @@ -699,7 +702,7 @@ void conn_event_do(const struct ctx *c, struct tcp_tap_conn *conn,
>  		tcp_epoll_ctl(c, conn);
>  
>  	if (CONN_HAS(conn, SOCK_FIN_SENT | TAP_FIN_ACKED))
> -		tcp_timer_ctl(c, conn);
> +		tcp_timer_ctl(conn);
>  }
>  
>  /**
> @@ -1737,7 +1740,7 @@ static int tcp_data_from_tap(const struct ctx *c, struct tcp_tap_conn *conn,
>  				   seq, conn->seq_from_tap);
>  
>  			tcp_send_flag(c, conn, ACK);
> -			tcp_timer_ctl(c, conn);
> +			tcp_timer_ctl(conn);
>  
>  			if (p->count == 1) {
>  				tcp_tap_window_update(c, conn,
> @@ -2386,7 +2389,7 @@ void tcp_timer_handler(const struct ctx *c, union epoll_ref ref)
>  
>  	if (conn->flags & ACK_TO_TAP_DUE) {
>  		tcp_send_flag(c, conn, ACK_IF_NEEDED);
> -		tcp_timer_ctl(c, conn);
> +		tcp_timer_ctl(conn);
>  	} else if (conn->flags & ACK_FROM_TAP_DUE) {
>  		if (!(conn->events & ESTABLISHED)) {
>  			flow_dbg(conn, "handshake timeout");
> @@ -2408,7 +2411,7 @@ void tcp_timer_handler(const struct ctx *c, union epoll_ref ref)
>  				return;
>  
>  			tcp_data_from_sock(c, conn);
> -			tcp_timer_ctl(c, conn);
> +			tcp_timer_ctl(conn);
>  		}
>  	} else {
>  		struct itimerspec new = { { 0 }, { ACT_TIMEOUT, 0 } };
> @@ -3456,7 +3459,7 @@ int tcp_flow_migrate_source_ext(const struct ctx *c,
>  	if (c->migrate_no_linger)
>  		close(s);
>  	else
> -		epoll_del(c->epollfd, s);
> +		epoll_del(flow_epollfd_get(&conn->f), s);
>  
>  	/* Adjustments unrelated to FIN segments: sequence numbers we dumped are
>  	 * based on the end of the queues.
> @@ -3605,7 +3608,7 @@ static int tcp_flow_repair_connect(const struct ctx *c,
>  		return rc;
>  	}
>  
> -	conn->in_epoll = 0;
> +	conn->f.threadnb = FLOW_THREADNB_INVALID;
>  	conn->timer = -1;
>  	conn->listening_sock = -1;
>  
> diff --git a/tcp_conn.h b/tcp_conn.h
> index 38b5c541f003..81333122d531 100644
> --- a/tcp_conn.h
> +++ b/tcp_conn.h
> @@ -12,7 +12,6 @@
>  /**
>   * struct tcp_tap_conn - Descriptor for a TCP connection (not spliced)
>   * @f:			Generic flow information
> - * @in_epoll:		Is the connection in the epoll set?
>   * @retrans:		Number of retransmissions occurred due to ACK_TIMEOUT
>   * @ws_from_tap:	Window scaling factor advertised from tap/guest
>   * @ws_to_tap:		Window scaling factor advertised to tap/guest
> @@ -36,8 +35,6 @@ struct tcp_tap_conn {
>  	/* Must be first element */
>  	struct flow_common f;
>  
> -	bool		in_epoll	:1;
> -

One wrinkle with moving this to flow_common is that we maybe need to
be a bit clearer about what the semantics of "in_epoll" vs. not
"in_epoll" are - however it's encoded.  After all some flow types have
multiple fds, which could independently be in or out of the epoll.

I think I considered moving in_epoll to flow_common before, but
decided against it, because I punted on pinning down those semantics:
as a per flow type flag, it has a per flow type meaning.

>  #define TCP_RETRANS_BITS		3
>  	unsigned int	retrans		:TCP_RETRANS_BITS;
>  #define TCP_MAX_RETRANS			MAX_FROM_BITS(TCP_RETRANS_BITS)
> @@ -196,7 +193,6 @@ struct tcp_tap_transfer_ext {
>   * @written:		Bytes written (not fully written from one other side read)
>   * @events:		Events observed/actions performed on connection
>   * @flags:		Connection flags (attributes, not events)
> - * @in_epoll:		Is the connection in the epoll set?
>   */
>  struct tcp_splice_conn {
>  	/* Must be first element */
> @@ -220,8 +216,6 @@ struct tcp_splice_conn {
>  #define RCVLOWAT_SET(sidei_)		((sidei_) ? BIT(1) : BIT(0))
>  #define RCVLOWAT_ACT(sidei_)		((sidei_) ? BIT(3) : BIT(2))
>  #define CLOSING				BIT(4)
> -
> -	bool in_epoll	:1;
>  };
>  
>  /* Socket pools */
> @@ -245,7 +239,7 @@ int tcp_flow_migrate_target_ext(struct ctx *c, struct tcp_tap_conn *conn, int fd
>  bool tcp_flow_is_established(const struct tcp_tap_conn *conn);
>  
>  bool tcp_splice_flow_defer(struct tcp_splice_conn *conn);
> -void tcp_splice_timer(const struct ctx *c, struct tcp_splice_conn *conn);
> +void tcp_splice_timer(struct tcp_splice_conn *conn);
>  int tcp_conn_pool_sock(int pool[]);
>  int tcp_conn_sock(sa_family_t af);
>  int tcp_sock_refill_pool(int pool[], sa_family_t af);
> diff --git a/tcp_splice.c b/tcp_splice.c
> index 6f21184bdc55..8842feef0684 100644
> --- a/tcp_splice.c
> +++ b/tcp_splice.c
> @@ -149,7 +149,9 @@ static void tcp_splice_conn_epoll_events(uint16_t events,
>  static int tcp_splice_epoll_ctl(const struct ctx *c,
>  				struct tcp_splice_conn *conn)
>  {
> -	int m = conn->in_epoll ? EPOLL_CTL_MOD : EPOLL_CTL_ADD;
> +	int epollfd = flow_epollfd_valid(&conn->f) ? flow_epollfd_get(&conn->f)
> +						   : c->epollfd;
> +	int m = flow_epollfd_valid(&conn->f) ? EPOLL_CTL_MOD : EPOLL_CTL_ADD;
>  	const union epoll_ref ref[SIDES] = {
>  		{ .type = EPOLL_TYPE_TCP_SPLICE, .fd = conn->s[0],
>  		  .flowside = FLOW_SIDX(conn, 0) },
> @@ -161,25 +163,24 @@ static int tcp_splice_epoll_ctl(const struct ctx *c,
>  
>  	tcp_splice_conn_epoll_events(conn->events, ev);
>  
> -	if (epoll_ctl(c->epollfd, m, conn->s[0], &ev[0]) ||
> -	    epoll_ctl(c->epollfd, m, conn->s[1], &ev[1])) {
> +
> +	if (epoll_ctl(epollfd, m, conn->s[0], &ev[0]) ||
> +	    epoll_ctl(epollfd, m, conn->s[1], &ev[1])) {
>  		int ret = -errno;
>  		flow_perror(conn, "ERROR on epoll_ctl()");
>  		return ret;
>  	}
> -
> -	conn->in_epoll = true;
> +	flow_epollfd_set(&conn->f, 0, epollfd);
>  
>  	return 0;
>  }
>  
>  /**
>   * conn_flag_do() - Set/unset given flag, log, update epoll on CLOSING flag
> - * @c:		Execution context
>   * @conn:	Connection pointer
>   * @flag:	Flag to set, or ~flag to unset
>   */
> -static void conn_flag_do(const struct ctx *c, struct tcp_splice_conn *conn,
> +static void conn_flag_do(struct tcp_splice_conn *conn,
>  			 unsigned long flag)
>  {
>  	if (flag & (flag - 1)) {
> @@ -204,15 +205,15 @@ static void conn_flag_do(const struct ctx *c, struct tcp_splice_conn *conn,
>  	}
>  
>  	if (flag == CLOSING) {
> -		epoll_del(c->epollfd, conn->s[0]);
> -		epoll_del(c->epollfd, conn->s[1]);
> +		epoll_del(flow_epollfd_get(&conn->f), conn->s[0]);
> +		epoll_del(flow_epollfd_get(&conn->f), conn->s[1]);
>  	}
>  }
>  
>  #define conn_flag(c, conn, flag)					\
>  	do {								\
>  		flow_trace(conn, "flag at %s:%i", __func__, __LINE__);	\
> -		conn_flag_do(c, conn, flag);				\
> +		conn_flag_do(conn, flag);				\
>  	} while (0)
>  
>  /**
> @@ -751,10 +752,9 @@ void tcp_splice_init(struct ctx *c)
>  
>  /**
>   * tcp_splice_timer() - Timer for spliced connections
> - * @c:		Execution context
>   * @conn:	Connection to handle
>   */
> -void tcp_splice_timer(const struct ctx *c, struct tcp_splice_conn *conn)
> +void tcp_splice_timer(struct tcp_splice_conn *conn)
>  {
>  	unsigned sidei;
>  
> -- 
> 2.50.1
> 

-- 
David Gibson (he or they)	| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you, not the other way
				| around.
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]