For each each direction of each spliced connection, we keep track of how many bytes we've read from one socket and written to the other. However, we never actually care about the absolute values of these, only the difference between them, which represents how much data is currently "in flight" in the splicing pipe. Simplify the handling by having a single variable tracking the number of bytes in the pipe. As a bonus, the new scheme makes it clearer that we don't need to worry about overflows: pending can never become larger than the maximum pipe bufffer size, well within 32-bits. I _think_ the old scheme was safe in the case of overflow - again under the assumption that read/written can never be further apart than the pipe buffer size. However, it's much harder to reason about this case. It's certainly plausible that an overflow could occur - sending 4GiB through a local socket is entirely achievable. Signed-off-by: David Gibson --- tcp_conn.h | 6 ++---- tcp_splice.c | 18 +++++++++--------- 2 files changed, 11 insertions(+), 13 deletions(-) diff --git a/tcp_conn.h b/tcp_conn.h index 9f5bee03..c8381aa7 100644 --- a/tcp_conn.h +++ b/tcp_conn.h @@ -206,8 +206,7 @@ struct tcp_tap_transfer_ext { * @f: Generic flow information * @s: File descriptor for sockets * @pipe: File descriptors for pipes - * @read: Bytes read (not fully written to other side in one shot) - * @written: Bytes written (not fully written from one other side read) + * @pending: Bytes currently in each pipe * @events: Events observed/actions performed on connection * @flags: Connection flags (attributes, not events) */ @@ -218,8 +217,7 @@ struct tcp_splice_conn { int s[SIDES]; int pipe[SIDES][2]; - uint32_t read[SIDES]; - uint32_t written[SIDES]; + uint32_t pending[SIDES]; uint8_t events; #define SPLICE_CLOSED 0 diff --git a/tcp_splice.c b/tcp_splice.c index 18e8b303..8fbd490f 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -292,7 +292,7 @@ bool tcp_splice_flow_defer(struct tcp_splice_conn *conn) conn->s[sidei] = -1; } - conn->read[sidei] = conn->written[sidei] = 0; + conn->pending[sidei] = 0; } conn->events = SPLICE_CLOSED; @@ -490,7 +490,7 @@ static int tcp_splice_forward(struct ctx *c, struct int eof = 0; while (1) { - ssize_t readlen, written, pending; + ssize_t readlen, written; int more = 0; retry: @@ -537,7 +537,7 @@ retry: flow_trace(conn, "%zi from write-side call (passed %zi)", written, c->tcp.pipe_size); - /* Most common case: skip updating counters. */ + /* Most common case: skip updating pending. */ if (readlen > 0 && readlen == written) { if (readlen >= (long)c->tcp.pipe_size * 10 / 100) continue; @@ -561,11 +561,11 @@ retry: continue; } - conn->read[fromsidei] += readlen > 0 ? readlen : 0; - conn->written[fromsidei] += written > 0 ? written : 0; + conn->pending[fromsidei] += readlen > 0 ? readlen : 0; + conn->pending[fromsidei] -= written > 0 ? written : 0; if (written < 0) { - if (conn->read[fromsidei] == conn->written[fromsidei]) + if (!conn->pending[fromsidei]) break; conn_event(conn, OUT_WAIT(!fromsidei)); @@ -575,15 +575,15 @@ retry: if (never_read && written == (long)(c->tcp.pipe_size)) goto retry; - pending = conn->read[fromsidei] - conn->written[fromsidei]; - if (!never_read && written > 0 && written < pending) + if (!never_read && written > 0 && + written < conn->pending[fromsidei]) goto retry; if (eof) break; } - if (conn->read[fromsidei] == conn->written[fromsidei] && eof) { + if (!conn->pending[fromsidei] && eof) { unsigned sidei; flow_foreach_sidei(sidei) { -- 2.54.0

There are two ways we can tell one of our sockets has received a FIN. We can either see an EPOLLRDHUP epoll event, or we can get a zero-length read (EOF) on the socket. We currently use both, in a mildly confusing way: we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then some other close out logic is based on seeing an EOF. Simplify this by setting the flag based on only the EOF. To make sure we don't miss an event if we get an EPOLLRDHUP with no data, we trigger the forwarding path for EPOLLRDHUP as well as EPOLLIN. Signed-off-by: David Gibson --- tcp_splice.c | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-) diff --git a/tcp_splice.c b/tcp_splice.c index 8fbd490f..b45f0060 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); int never_read = 1; - int eof = 0; while (1) { ssize_t readlen, written; @@ -510,7 +509,7 @@ retry: flow_trace(conn, "%zi from read-side call", readlen); if (!readlen) { - eof = 1; + conn_event(conn, FIN_RCVD(fromsidei)); } else if (readlen > 0) { never_read = 0; @@ -579,11 +578,12 @@ retry: written < conn->pending[fromsidei]) goto retry; - if (eof) + if (conn->events & FIN_RCVD(fromsidei)) break; } - if (!conn->pending[fromsidei] && eof) { + if (!conn->pending[fromsidei] && + conn->events & FIN_RCVD(fromsidei)) { unsigned sidei; flow_foreach_sidei(sidei) { @@ -643,17 +643,13 @@ void tcp_splice_sock_handler(struct ctx *c, union epoll_ref ref, goto reset; } - if (events & EPOLLRDHUP) - /* For side 0 this is fake, but implied */ - conn_event(conn, FIN_RCVD(evsidei)); - if (events & EPOLLOUT) { if (tcp_splice_forward(c, conn, !evsidei)) goto reset; conn_event(conn, ~OUT_WAIT(evsidei)); } - if (events & EPOLLIN) { + if (events & (EPOLLIN | EPOLLRDHUP)) { if (tcp_splice_forward(c, conn, evsidei)) goto reset; } -- 2.54.0

On Wed, 20 May 2026 23:08:47 +1000 David Gibson wrote:

tcp_splice_sock_handler() has an optimised path for the common case where the amount we splice(2) into the pipe is exactly the same as the amount we splice(2) out again. If the pipe is empty at that point, we stop forwarding until we get another epoll event.

However, via a subtle chain of events, this can cause a bug for a half-closed connection. Suppose the connection is already half-closed in the other direction - that is, we've already called shutdown(SHUT_WR) on the socket for which we're getting the event. In this event we're getting the last batch of data in the other direction, and also a FIN. This can result in EPOLLIN, EPOLLRDHUP and EPOLLHUP events simultaneously.

We read the last data from the socket and successfully splice it to the other side. Since there is no data in the pipe, we exit the forwarding loop. However, because we did read data, we don't set the eof flag.

Because we don't set eof, we don't (yet) propagate the FIN to the other side, or set FIN_SENT_(!fromsidei). Therefore we don't (yet) recognize this as a clean termination and set the CLOSING flag. We would correct this when we get our next event, however before we can do so we process the EPOLLHUP event. Because we haven't recognized this as a clean close we assume it is an abrupt close and send an RST to the other side.

To avoid this, don't stop attempting to forward data on this path. Continue for at least one more loop. If we're at EOF, we'll recognize it on the next splice(2). If not it gives us an opportunity to forward more data without returning to the mail epoll loop.

Oops. The fix looks correct to me, but I wonder: is it clear to you why the issue only started occurring in this release? This code had "always" been there. I see a few possible directions but I'm not quite sure. Not that important anyway, if you could reproduce the issue and this fixes it. Just one nit:

Link: https://bugs.passt.top/show_bug.cgi?id=202 Signed-off-by: David Gibson

Reported-by: Paul Holzinger

--- tcp_splice.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tcp_splice.c b/tcp_splice.c index 1359d6b8..34ffea73 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -605,7 +605,7 @@ retry: } }

- break; + continue; }

conn->read[fromsidei] += readlen > 0 ? readlen : 0;

-- Stefano

On Wed, 20 May 2026 23:08:50 +1000 David Gibson wrote:

There are two ways we can tell one of our sockets has received a FIN. We can either see an EPOLLRDHUP epoll event, or we can get a zero-length read (EOF) on the socket. We currently use both, in a mildly confusing way: we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then some other close out logic is based on seeing an EOF.

Simplify this by setting the flag based on only the EOF. To make sure we don't miss an event if we get an EPOLLRDHUP with no data, we trigger the forwarding path for EPOLLRDHUP as well as EPOLLIN.

Signed-off-by: David Gibson --- tcp_splice.c | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/tcp_splice.c b/tcp_splice.c index 8fbd490f..b45f0060 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); int never_read = 1; - int eof = 0;

while (1) { ssize_t readlen, written; @@ -510,7 +509,7 @@ retry: flow_trace(conn, "%zi from read-side call", readlen);

if (!readlen) { - eof = 1; + conn_event(conn, FIN_RCVD(fromsidei));

I'm not sure if I really found a concrete issue with this, but it looks a bit scary, because it changes the semantics of FIN_RCVD, which used to mean that we infer we received a FIN, regardless of whether we're done processing all data from that half of the connection. Now FIN_RCVD is only set if we actually processed all the data and we hit the end of file. The (potential) issue I see here is that we get EPOLLRDHUP, splice() returns -1 with EAGAIN in errno because we had no room in the pipe, and it would have returned 0 instead. Will we ever get our zero-sized "read" later? If not, we might have missed EPOLLRDHUP *and* the end of file. I'm not entirely sure we have guarantees in that sense from splice(). The existing implementation distinguishes between end-of-file we hit in a given iteration, and EPOLLRDHUP we might have seen at any time. That was actually intended. -- Stefano

On Wed, May 20, 2026 at 10:30:04PM +0200, Stefano Brivio wrote:

On Wed, 20 May 2026 23:08:50 +1000 David Gibson wrote:

...

There are two ways we can tell one of our sockets has received a FIN. We can either see an EPOLLRDHUP epoll event, or we can get a zero-length read (EOF) on the socket. We currently use both, in a mildly confusing way: we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then some other close out logic is based on seeing an EOF.

Simplify this by setting the flag based on only the EOF. To make sure we don't miss an event if we get an EPOLLRDHUP with no data, we trigger the forwarding path for EPOLLRDHUP as well as EPOLLIN.

Signed-off-by: David Gibson --- tcp_splice.c | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/tcp_splice.c b/tcp_splice.c index 8fbd490f..b45f0060 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); int never_read = 1; - int eof = 0;

while (1) { ssize_t readlen, written; @@ -510,7 +509,7 @@ retry: flow_trace(conn, "%zi from read-side call", readlen);

if (!readlen) { - eof = 1; + conn_event(conn, FIN_RCVD(fromsidei));

I'm not sure if I really found a concrete issue with this, but it looks a bit scary, because it changes the semantics of FIN_RCVD, which used to mean that we infer we received a FIN, regardless of whether we're done processing all data from that half of the connection.

Now FIN_RCVD is only set if we actually processed all the data and we hit the end of file.

True. But the only place that tested FIN_RCVD was at the end of tcp_splice_forward(), conditional on 'eof' anyway. In a sense, this was the cause of bug202 - we had FIN_RCVD set, but we didn't process it and shutdown() on the other side, because we didn't have eof.

The (potential) issue I see here is that we get EPOLLRDHUP, splice() returns -1 with EAGAIN in errno because we had no room in the pipe, and it would have returned 0 instead.

Will we ever get our zero-sized "read" later? If not, we might have missed EPOLLRDHUP *and* the end of file. I'm not entirely sure we have guarantees in that sense from splice().

It's not really about guarantees from splice. I'm pretty sure this is ok, reasoning as follows. Consider all the exit points from the loop body: - Each return is a return -1, so we kill the connection anyway. They don't matter - Each continue, goto retry and the end of the body will do the read side splice() again, so get another chance to see the EOF - That leaves just the breaks Consider each break (there are three, since patch 2 of this series) if (written < 0) { if (!conn->pending[fromsidei]) break; (1) The pipe is empty and the write-splice returned EAGAIN, so it didn't remove data from the pipe. Therefore, the pipe must have been empty before the write-splice. Which means the read-splice can't have blocked on a full pipe. conn_event(conn, OUT_WAIT(!fromsidei)); break; } (2) The pipe is non-empty and the write-splice returned EAGAIN, so it must have blocked on the output socket. We've set OUT_WAIT(), so we'll get an EPOLLOUT at some point which will cause us to read-splice again, meaning we get another chance to see the EOF. [...] if (conn->events & FIN_RCVD(fromsidei)) break; (3) By the new semantics of FIN_RCVD, we *have* seen the EOF.

The existing implementation distinguishes between end-of-file we hit in a given iteration, and EPOLLRDHUP we might have seen at any time. That was actually intended.

It might be intended, but I can't see that we did anything with that information. That said the conditions on which we exit / retry this loop are pretty darn confusing. I'll see if I can improve them. -- 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

On Wed, May 20, 2026 at 10:29:12PM +0200, Stefano Brivio wrote:

On Wed, 20 May 2026 23:08:49 +1000 David Gibson wrote:

...

For each each direction of each spliced connection, we keep track of how many bytes we've read from one socket and written to the other. However, we never actually care about the absolute values of these, only the difference between them, which represents how much data is currently "in flight" in the splicing pipe.

Simplify the handling by having a single variable tracking the number of bytes in the pipe.

For me it actually looks slightly more complicated to think about it this way, I added explicit 'read' and 'written' after being bitten by some issue I introduced with a previous 'pending' concept, but I have to admit it slightly simplifies the overflow topic.

...

As a bonus, the new scheme makes it clearer that we don't need to worry about overflows: pending can never become larger than the maximum pipe bufffer size, well within 32-bits.

I _think_ the old scheme was safe in the case of overflow - again under the assumption that read/written can never be further apart than the pipe buffer size. However, it's much harder to reason about this case. It's certainly plausible that an overflow could occur - sending 4GiB through a local socket is entirely achievable.

For me it looked pretty simple: you can overflow 32 bits (at 100 Gbps, but without hitting the "optimised" case, it would take about five minutes), but all the operations between the two counters are between two uint32_t, so they would happen in uint32_t, hence modulo 32 bits, similar to TCP sequences.

Plus it's only equality comparisons, so we don't need SEQ_GT or the like. Yeah, that's the reasoning, but to me that's still a lot more than "can't exceed pipe size".

Anyway, overall, I think it's an improvement over the original. One nit here:

...

Signed-off-by: David Gibson --- tcp_conn.h | 6 ++---- tcp_splice.c | 18 +++++++++--------- 2 files changed, 11 insertions(+), 13 deletions(-)

diff --git a/tcp_conn.h b/tcp_conn.h index 9f5bee03..c8381aa7 100644 --- a/tcp_conn.h +++ b/tcp_conn.h @@ -206,8 +206,7 @@ struct tcp_tap_transfer_ext { * @f: Generic flow information * @s: File descriptor for sockets * @pipe: File descriptors for pipes - * @read: Bytes read (not fully written to other side in one shot) - * @written: Bytes written (not fully written from one other side read) + * @pending: Bytes currently in each pipe * @events: Events observed/actions performed on connection * @flags: Connection flags (attributes, not events) */ @@ -218,8 +217,7 @@ struct tcp_splice_conn { int s[SIDES]; int pipe[SIDES][2];

- uint32_t read[SIDES]; - uint32_t written[SIDES]; + uint32_t pending[SIDES];

uint8_t events; #define SPLICE_CLOSED 0 diff --git a/tcp_splice.c b/tcp_splice.c index 18e8b303..8fbd490f 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -292,7 +292,7 @@ bool tcp_splice_flow_defer(struct tcp_splice_conn *conn) conn->s[sidei] = -1; }

- conn->read[sidei] = conn->written[sidei] = 0; + conn->pending[sidei] = 0; }

conn->events = SPLICE_CLOSED; @@ -490,7 +490,7 @@ static int tcp_splice_forward(struct ctx *c, struct int eof = 0;

while (1) { - ssize_t readlen, written, pending; + ssize_t readlen, written; int more = 0;

retry: @@ -537,7 +537,7 @@ retry: flow_trace(conn, "%zi from write-side call (passed %zi)", written, c->tcp.pipe_size);

- /* Most common case: skip updating counters. */ + /* Most common case: skip updating pending. */

"pending" isn't a noun (even though the variable name is, but it's not quite obvious that you're referring to it). I think that:

/* Most common case: skip updating count of pending bytes */

would be slightly clearer (and also omit the '.' because it's not a complete sentence, as we usually do on single-line comments, similarly to most occurrences in the kernel).

Done.

...

if (readlen > 0 && readlen == written) { if (readlen >= (long)c->tcp.pipe_size * 10 / 100) continue; @@ -561,11 +561,11 @@ retry: continue; }

- conn->read[fromsidei] += readlen > 0 ? readlen : 0; - conn->written[fromsidei] += written > 0 ? written : 0; + conn->pending[fromsidei] += readlen > 0 ? readlen : 0; + conn->pending[fromsidei] -= written > 0 ? written : 0;

if (written < 0) { - if (conn->read[fromsidei] == conn->written[fromsidei]) + if (!conn->pending[fromsidei]) break;

conn_event(conn, OUT_WAIT(!fromsidei)); @@ -575,15 +575,15 @@ retry: if (never_read && written == (long)(c->tcp.pipe_size)) goto retry;

- pending = conn->read[fromsidei] - conn->written[fromsidei]; - if (!never_read && written > 0 && written < pending) + if (!never_read && written > 0 && + written < conn->pending[fromsidei]) goto retry;

if (eof) break; }

- if (conn->read[fromsidei] == conn->written[fromsidei] && eof) { + if (!conn->pending[fromsidei] && eof) { unsigned sidei;

flow_foreach_sidei(sidei) {

-- Stefano

-- 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

On Thu, 21 May 2026 12:11:49 +1000 David Gibson wrote:

On Wed, May 20, 2026 at 10:30:23PM +0200, Stefano Brivio wrote:

...

On Wed, 20 May 2026 23:08:51 +1000 David Gibson wrote:

...

At the end of tcp_splice_forward(), we check for half-closed connections and propagate the FIN to the other side with a shutdown(2). Currently we check for a half closed connection in either direction. That's unnecessary here, because tcp_splice_forward() will already be called for each direction if there are any relevant events.

True, but do we have the guarantee that tcp_splice_forward() will also be called once all relevant FIN_RCVD / FIN_SENT flags have been sent?

Yes, because tcp_splice_forward() is (now) the only place that *sets* FIN_RCVD (or FIN_SENT).

...

The reason why we check both sides here is that we might have updated flags for one side, and now we need to double check if it's time to call shutdown() as a consequence.

Maybe we never have to, but I think it's not really obvious to prove.

tcp_splice_forward() only touches FIN_RCVD(fromsidei). So, we only need to examine FIN_RCVD(fromsidei). If FIN_RCVD is set for the other side, it must be in another call to tcp_splice_forward() which will also examine that other flag and shutdown() as necessary.

Okay, fair. That's what I call "not really obvious" but it's probably obvious enough.

...

...

Signed-off-by: David Gibson --- tcp_splice.c | 22 ++++++++-------------- 1 file changed, 8 insertions(+), 14 deletions(-)

diff --git a/tcp_splice.c b/tcp_splice.c index b45f0060..e5018f2e 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -582,21 +582,15 @@ retry: break; }

- if (!conn->pending[fromsidei] && - conn->events & FIN_RCVD(fromsidei)) { - unsigned sidei; - - flow_foreach_sidei(sidei) { - if ((conn->events & FIN_RCVD(sidei)) && - !(conn->events & FIN_SENT(!sidei))) { - if (shutdown(conn->s[!sidei], SHUT_WR) < 0) { - flow_perror(conn, "shutdown() on %s", - pif_name(conn->f.pif[!sidei])); - return -1; - } - conn_event(conn, FIN_SENT(!sidei)); - } + if ((conn->events & FIN_RCVD(fromsidei)) && + !(conn->events & FIN_SENT(!fromsidei)) && + !conn->pending[fromsidei]) { + if (shutdown(conn->s[!fromsidei], SHUT_WR) < 0) { + flow_perror(conn, "shutdown() on %s", + pif_name(conn->f.pif[!fromsidei])); + return -1; } + conn_event(conn, FIN_SENT(!fromsidei)); }

return 0;

-- Stefano

On Thu, May 21, 2026 at 07:40:31AM +0200, Stefano Brivio wrote:

On Thu, 21 May 2026 12:03:33 +1000 David Gibson wrote:

...

On Wed, May 20, 2026 at 10:30:04PM +0200, Stefano Brivio wrote:

...

On Wed, 20 May 2026 23:08:50 +1000 David Gibson wrote:

...

There are two ways we can tell one of our sockets has received a FIN. We can either see an EPOLLRDHUP epoll event, or we can get a zero-length read (EOF) on the socket. We currently use both, in a mildly confusing way: we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then some other close out logic is based on seeing an EOF.

Simplify this by setting the flag based on only the EOF. To make sure we don't miss an event if we get an EPOLLRDHUP with no data, we trigger the forwarding path for EPOLLRDHUP as well as EPOLLIN.

Signed-off-by: David Gibson --- tcp_splice.c | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/tcp_splice.c b/tcp_splice.c index 8fbd490f..b45f0060 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); int never_read = 1; - int eof = 0;

while (1) { ssize_t readlen, written; @@ -510,7 +509,7 @@ retry: flow_trace(conn, "%zi from read-side call", readlen);

if (!readlen) { - eof = 1; + conn_event(conn, FIN_RCVD(fromsidei));

I'm not sure if I really found a concrete issue with this, but it looks a bit scary, because it changes the semantics of FIN_RCVD, which used to mean that we infer we received a FIN, regardless of whether we're done processing all data from that half of the connection.

Now FIN_RCVD is only set if we actually processed all the data and we hit the end of file.

True. But the only place that tested FIN_RCVD was at the end of tcp_splice_forward(), conditional on 'eof' anyway. In a sense, this was the cause of bug202 - we had FIN_RCVD set, but we didn't process it and shutdown() on the other side, because we didn't have eof.

That sounds like a good motivation to clean this up, just two concerns below:

...

...

The (potential) issue I see here is that we get EPOLLRDHUP, splice() returns -1 with EAGAIN in errno because we had no room in the pipe, and it would have returned 0 instead.

Will we ever get our zero-sized "read" later? If not, we might have missed EPOLLRDHUP *and* the end of file. I'm not entirely sure we have guarantees in that sense from splice().

It's not really about guarantees from splice. I'm pretty sure this is ok, reasoning as follows.

Consider all the exit points from the loop body: - Each return is a return -1, so we kill the connection anyway. They don't matter - Each continue, goto retry and the end of the body will do the read side splice() again, so get another chance to see the EOF - That leaves just the breaks

Consider each break (there are three, since patch 2 of this series) if (written < 0) { if (!conn->pending[fromsidei]) break;

(1) The pipe is empty and the write-splice returned EAGAIN, so it didn't remove data from the pipe.

You're assuming that !conn->pending[fromsidei] means that the pipe is empty. From what we see of it, it is.

It does mean the pipe is empty. Everything we put in, we've taken out. There cannot be anything in there.

What the kernel can do with it, though, is different. It might return EAGAIN even if we think we should have space, because it's resizing it under memory pressure or anything like that. Or it delays freeing up space or accounting for whatever reason.

Theoretically, I suppose. But !pending doesn't just mean the pipe is not full it means it's completely completely empty. Not being able to put any bytes at all into an empty pipe would be *very* surprising. So much so that if it happened in practice, I suspect we wouldn't be safe not having epoll events on the pipe ends, so that we can be notified when it deigns to accept some data.

So it would be nice to make this part robust to that. I thought setting FIN_RCVD on EPOLLRDHUP was a good way to achieve that.

...

Therefore, the pipe must have been empty before the write-splice. Which means the read-splice can't have blocked on a full pipe. conn_event(conn, OUT_WAIT(!fromsidei)); break; }

(2) The pipe is non-empty and the write-splice returned EAGAIN, so it must have blocked on the output socket. We've set OUT_WAIT(), so we'll get an EPOLLOUT at some point which will cause us to read-splice again, meaning we get another chance to see the EOF.

...later. But what if we don't get a zero-sized read *at all*? I'm not sure if splice() guarantees we do get one if we reach end-of-file.

That's something valid and very well established for read() and recv(), but splice() is a bit weird. The documentation says:

A return value of 0 means end of input.

but I wouldn't assume we'll *always* get at least one in case of EOF.

What else could we plausibly get?

...

[...] if (conn->events & FIN_RCVD(fromsidei)) break; (3) By the new semantics of FIN_RCVD, we *have* seen the EOF.

...

The existing implementation distinguishes between end-of-file we hit in a given iteration, and EPOLLRDHUP we might have seen at any time. That was actually intended.

It might be intended, but I can't see that we did anything with that information.

We always set FIN_RCVD on it. You're right, if we only checked that on 'eof', that didn't solve much, but that wasn't necessarily intended. My original intention was to make setting of FIN_RCVD (or whatever it was originally) robust.

Ok, well. I've spotted other changes to make in the vicinity that I think will make some of this easier to reason about anyway. So I'll consider your points as I rework this and other patches.

...

That said the conditions on which we exit / retry this loop are pretty darn confusing. I'll see if I can improve them.

-- Stefano

-- 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

On Thu, May 21, 2026 at 09:15:13AM +0200, Stefano Brivio wrote:

On Thu, 21 May 2026 16:56:43 +1000 David Gibson wrote:

...

On Thu, May 21, 2026 at 07:40:31AM +0200, Stefano Brivio wrote:

...

On Thu, 21 May 2026 12:03:33 +1000 David Gibson wrote:

...

On Wed, May 20, 2026 at 10:30:04PM +0200, Stefano Brivio wrote:

...

On Wed, 20 May 2026 23:08:50 +1000 David Gibson wrote:

...

There are two ways we can tell one of our sockets has received a FIN. We can either see an EPOLLRDHUP epoll event, or we can get a zero-length read (EOF) on the socket. We currently use both, in a mildly confusing way: we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then some other close out logic is based on seeing an EOF.

Simplify this by setting the flag based on only the EOF. To make sure we don't miss an event if we get an EPOLLRDHUP with no data, we trigger the forwarding path for EPOLLRDHUP as well as EPOLLIN.

Signed-off-by: David Gibson --- tcp_splice.c | 14 +++++--------- 1 file changed, 5 insertions(+), 9 deletions(-)

diff --git a/tcp_splice.c b/tcp_splice.c index 8fbd490f..b45f0060 100644 --- a/tcp_splice.c +++ b/tcp_splice.c @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); int never_read = 1; - int eof = 0;

while (1) { ssize_t readlen, written; @@ -510,7 +509,7 @@ retry: flow_trace(conn, "%zi from read-side call", readlen);

if (!readlen) { - eof = 1; + conn_event(conn, FIN_RCVD(fromsidei));

I'm not sure if I really found a concrete issue with this, but it looks a bit scary, because it changes the semantics of FIN_RCVD, which used to mean that we infer we received a FIN, regardless of whether we're done processing all data from that half of the connection.

Now FIN_RCVD is only set if we actually processed all the data and we hit the end of file.

True. But the only place that tested FIN_RCVD was at the end of tcp_splice_forward(), conditional on 'eof' anyway. In a sense, this was the cause of bug202 - we had FIN_RCVD set, but we didn't process it and shutdown() on the other side, because we didn't have eof.

That sounds like a good motivation to clean this up, just two concerns below:

...

...

The (potential) issue I see here is that we get EPOLLRDHUP, splice() returns -1 with EAGAIN in errno because we had no room in the pipe, and it would have returned 0 instead.

Will we ever get our zero-sized "read" later? If not, we might have missed EPOLLRDHUP *and* the end of file. I'm not entirely sure we have guarantees in that sense from splice().

It's not really about guarantees from splice. I'm pretty sure this is ok, reasoning as follows.

Consider all the exit points from the loop body: - Each return is a return -1, so we kill the connection anyway. They don't matter - Each continue, goto retry and the end of the body will do the read side splice() again, so get another chance to see the EOF - That leaves just the breaks

Consider each break (there are three, since patch 2 of this series) if (written < 0) { if (!conn->pending[fromsidei]) break;

(1) The pipe is empty and the write-splice returned EAGAIN, so it didn't remove data from the pipe.

You're assuming that !conn->pending[fromsidei] means that the pipe is empty. From what we see of it, it is.

It does mean the pipe is empty. Everything we put in, we've taken out. There cannot be anything in there.

...

What the kernel can do with it, though, is different. It might return EAGAIN even if we think we should have space, because it's resizing it under memory pressure or anything like that. Or it delays freeing up space or accounting for whatever reason.

Theoretically, I suppose. But !pending doesn't just mean the pipe is not full it means it's completely completely empty. Not being able to put any bytes at all into an empty pipe would be *very* surprising. So much so that if it happened in practice, I suspect we wouldn't be safe not having epoll events on the pipe ends, so that we can be notified when it deigns to accept some data.

We can get 512-byte pipes, I actually saw that happening in practice with either:

Sure.. so? We can still put some bytes into it if it's empty.

- people setting low values for ulimits

- the user (or just pasta itself) having a lot of pipes open

and if I recall correctly that's where I saw the case of a supposedly empty pipe giving us EAGAIN. That was years ago though and I didn't specifically fix that.

I mean.. that sounds like a kernel bug. If we do have to handle that case we'll need epoll events on the pipe ends, since none of the socket events we monitor will trigger when the pipe becomes writable.

We currently probe the size based on the value we can have for 32 pipes (TCP_SPLICE_PIPE_POOL_SIZE). By making that 4096 or so you should get rather small pipes.

Things might already be broken with them, I haven't checked the behaviour in a long while. I think 512 bytes was the lower bound I hit.

...

...

So it would be nice to make this part robust to that. I thought setting FIN_RCVD on EPOLLRDHUP was a good way to achieve that.

...

Therefore, the pipe must have been empty before the write-splice. Which means the read-splice can't have blocked on a full pipe. conn_event(conn, OUT_WAIT(!fromsidei)); break; }

(2) The pipe is non-empty and the write-splice returned EAGAIN, so it must have blocked on the output socket. We've set OUT_WAIT(), so we'll get an EPOLLOUT at some point which will cause us to read-splice again, meaning we get another chance to see the EOF.

...later. But what if we don't get a zero-sized read *at all*? I'm not sure if splice() guarantees we do get one if we reach end-of-file.

...

That's something valid and very well established for read() and recv(), but splice() is a bit weird. The documentation says:

A return value of 0 means end of input.

but I wouldn't assume we'll *always* get at least one in case of EOF.

What else could we plausibly get?

-1 with EBADF, probably with EPOLLERR, because something timed out?

EBADF makes no sense, the fds are still valid, even if they're at EOF.

But I guess you're right, as long as we're not in the EPOLLERR category of things, we should consistently get 0, even if we read multiple times.

I had in mind some kernel behaviour where we get 0 once, and then -1 (EAGAIN?) because... go figure. But no, it can't happen.

I think the logic should be ok as long as we see a 0 once, even if we get EAGAINs after that. Another way to look at this - if we had to monitor EPOLLRDHUP to get this right, splice() would be unusable from blocking / synchronous code, which I don't think is the case.

...

...

...

[...] if (conn->events & FIN_RCVD(fromsidei)) break; (3) By the new semantics of FIN_RCVD, we *have* seen the EOF.

...

The existing implementation distinguishes between end-of-file we hit in a given iteration, and EPOLLRDHUP we might have seen at any time. That was actually intended.

It might be intended, but I can't see that we did anything with that information.

We always set FIN_RCVD on it. You're right, if we only checked that on 'eof', that didn't solve much, but that wasn't necessarily intended. My original intention was to make setting of FIN_RCVD (or whatever it was originally) robust.

Ok, well. I've spotted other changes to make in the vicinity that I think will make some of this easier to reason about anyway. So I'll consider your points as I rework this and other patches.

...

...

That said the conditions on which we exit / retry this loop are pretty darn confusing. I'll see if I can improve them.

-- Stefano

-- Stefano

-- 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

On Thu, May 21, 2026 at 05:18:16PM +0200, Stefano Brivio wrote:

On Thu, 21 May 2026 23:51:04 +1000 David Gibson wrote:

...

On Thu, May 21, 2026 at 09:15:13AM +0200, Stefano Brivio wrote:

...

On Thu, 21 May 2026 16:56:43 +1000 David Gibson wrote:

...

On Thu, May 21, 2026 at 07:40:31AM +0200, Stefano Brivio wrote:

...

On Thu, 21 May 2026 12:03:33 +1000 David Gibson wrote:

...

On Wed, May 20, 2026 at 10:30:04PM +0200, Stefano Brivio wrote: > On Wed, 20 May 2026 23:08:50 +1000 > David Gibson wrote: > > > There are two ways we can tell one of our sockets has received a FIN. We > > can either see an EPOLLRDHUP epoll event, or we can get a zero-length read > > (EOF) on the socket. We currently use both, in a mildly confusing way: > > we only set the FIN_RCVD() flag based on the EPOLLRDHUP event, but then > > some other close out logic is based on seeing an EOF. > > > > Simplify this by setting the flag based on only the EOF. To make sure we > > don't miss an event if we get an EPOLLRDHUP with no data, we trigger the > > forwarding path for EPOLLRDHUP as well as EPOLLIN. > > > > Signed-off-by: David Gibson > > --- > > tcp_splice.c | 14 +++++--------- > > 1 file changed, 5 insertions(+), 9 deletions(-) > > > > diff --git a/tcp_splice.c b/tcp_splice.c > > index 8fbd490f..b45f0060 100644 > > --- a/tcp_splice.c > > +++ b/tcp_splice.c > > @@ -487,7 +487,6 @@ static int tcp_splice_forward(struct ctx *c, struct > > uint8_t lowat_set_flag = RCVLOWAT_SET(fromsidei); > > uint8_t lowat_act_flag = RCVLOWAT_ACT(fromsidei); > > int never_read = 1; > > - int eof = 0; > > > > while (1) { > > ssize_t readlen, written; > > @@ -510,7 +509,7 @@ retry: > > flow_trace(conn, "%zi from read-side call", readlen); > > > > if (!readlen) { > > - eof = 1; > > + conn_event(conn, FIN_RCVD(fromsidei)); > > I'm not sure if I really found a concrete issue with this, but it looks > a bit scary, because it changes the semantics of FIN_RCVD, which used to > mean that we infer we received a FIN, regardless of whether we're done > processing all data from that half of the connection. > > Now FIN_RCVD is only set if we actually processed all the data and we > hit the end of file.

True. But the only place that tested FIN_RCVD was at the end of tcp_splice_forward(), conditional on 'eof' anyway. In a sense, this was the cause of bug202 - we had FIN_RCVD set, but we didn't process it and shutdown() on the other side, because we didn't have eof.

That sounds like a good motivation to clean this up, just two concerns below:

...

> The (potential) issue I see here is that we get EPOLLRDHUP, splice() > returns -1 with EAGAIN in errno because we had no room in the pipe, > and it would have returned 0 instead. > > Will we ever get our zero-sized "read" later? If not, we might have > missed EPOLLRDHUP *and* the end of file. I'm not entirely sure we have > guarantees in that sense from splice().

It's not really about guarantees from splice. I'm pretty sure this is ok, reasoning as follows.

Consider all the exit points from the loop body: - Each return is a return -1, so we kill the connection anyway. They don't matter - Each continue, goto retry and the end of the body will do the read side splice() again, so get another chance to see the EOF - That leaves just the breaks

Consider each break (there are three, since patch 2 of this series) if (written < 0) { if (!conn->pending[fromsidei]) break;

(1) The pipe is empty and the write-splice returned EAGAIN, so it didn't remove data from the pipe.

You're assuming that !conn->pending[fromsidei] means that the pipe is empty. From what we see of it, it is.

It does mean the pipe is empty. Everything we put in, we've taken out. There cannot be anything in there.

...

What the kernel can do with it, though, is different. It might return EAGAIN even if we think we should have space, because it's resizing it under memory pressure or anything like that. Or it delays freeing up space or accounting for whatever reason.

Theoretically, I suppose. But !pending doesn't just mean the pipe is not full it means it's completely completely empty. Not being able to put any bytes at all into an empty pipe would be *very* surprising. So much so that if it happened in practice, I suspect we wouldn't be safe not having epoll events on the pipe ends, so that we can be notified when it deigns to accept some data.

We can get 512-byte pipes, I actually saw that happening in practice with either:

Sure.. so? We can still put some bytes into it if it's empty.

The difference between empty and full is pretty small in that case.

Small, but not nothing. The empty pipe can still accept *some* bytes - otherwise the pipe is useless. Accept any bytes and the reasoning above works.

...

...

- people setting low values for ulimits

- the user (or just pasta itself) having a lot of pipes open

and if I recall correctly that's where I saw the case of a supposedly empty pipe giving us EAGAIN. That was years ago though and I didn't specifically fix that.

I mean.. that sounds like a kernel bug.

fcntl(2) says, for F_SETPIPE_SZ:

Note that because of the way the pages of the pipe buffer are em‐ ployed when data is written to the pipe, the number of bytes that can be written may be less than the nominal size, depending on the size of the writes.

...which I kind of understand really.

Ok, but still if an empty pipe can't accept *any* bytes, it's useless, which would make that a kernel bug.

...

If we do have to handle that case we'll need epoll events on the pipe ends, since none of the socket events we monitor will trigger when the pipe becomes writable.

Well, EPOLLOUT should do it.

But EPOLLOUT on the pipe end itself, not just on the write side socket as we have now.

...

...

We currently probe the size based on the value we can have for 32 pipes (TCP_SPLICE_PIPE_POOL_SIZE). By making that 4096 or so you should get rather small pipes.

Things might already be broken with them, I haven't checked the behaviour in a long while. I think 512 bytes was the lower bound I hit.

...

...

So it would be nice to make this part robust to that. I thought setting FIN_RCVD on EPOLLRDHUP was a good way to achieve that.

...

Therefore, the pipe must have been empty before the write-splice. Which means the read-splice can't have blocked on a full pipe. conn_event(conn, OUT_WAIT(!fromsidei)); break; }

(2) The pipe is non-empty and the write-splice returned EAGAIN, so it must have blocked on the output socket. We've set OUT_WAIT(), so we'll get an EPOLLOUT at some point which will cause us to read-splice again, meaning we get another chance to see the EOF.

...later. But what if we don't get a zero-sized read *at all*? I'm not sure if splice() guarantees we do get one if we reach end-of-file.

...

That's something valid and very well established for read() and recv(), but splice() is a bit weird. The documentation says:

A return value of 0 means end of input.

but I wouldn't assume we'll *always* get at least one in case of EOF.

What else could we plausibly get?

-1 with EBADF, probably with EPOLLERR, because something timed out?

EBADF makes no sense, the fds are still valid, even if they're at EOF.

I was thinking we hit EOF, don't notice right away, but after seconds / minutes and the socket is already closed.

The only place we close the socket is in the flow close path, at which point we've already decided it's over so the whole question is moot.

...

...

But I guess you're right, as long as we're not in the EPOLLERR category of things, we should consistently get 0, even if we read multiple times.

I had in mind some kernel behaviour where we get 0 once, and then -1 (EAGAIN?) because... go figure. But no, it can't happen.

I think the logic should be ok as long as we see a 0 once, even if we get EAGAINs after that.

Another way to look at this - if we had to monitor EPOLLRDHUP to get this right, splice() would be unusable from blocking / synchronous code, which I don't think is the case.

Right, yes, I'm fairly convinced at this point.

Ok :).

...

...

...

...

...

[...] if (conn->events & FIN_RCVD(fromsidei)) break; (3) By the new semantics of FIN_RCVD, we *have* seen the EOF.

> The existing implementation distinguishes between end-of-file we hit in > a given iteration, and EPOLLRDHUP we might have seen at any time. > That was actually intended.

It might be intended, but I can't see that we did anything with that information.

We always set FIN_RCVD on it. You're right, if we only checked that on 'eof', that didn't solve much, but that wasn't necessarily intended. My original intention was to make setting of FIN_RCVD (or whatever it was originally) robust.

Ok, well. I've spotted other changes to make in the vicinity that I think will make some of this easier to reason about anyway. So I'll consider your points as I rework this and other patches.

...

...

That said the conditions on which we exit / retry this loop are pretty darn confusing. I'll see if I can improve them.

-- Stefano

-- 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