kernel/linux.git/include/net/tcp.h, branch linux-6.5.y

tcp: fix cookie_init_timestamp() overflows

2023-11-20T10:56:52+00:00

[ Upstream commit 73ed8e03388d16c12fc577e5c700b58a29045a15 ] cookie_init_timestamp() is supposed to return a 64bit timestamp suitable for both TSval determination and setting of skb->tstamp. Unfortunately it uses 32bit fields and overflows after 2^32 * 10^6 nsec (~49 days) of uptime. Generated TSval are still correct, but skb->tstamp might be set far away in the past, potentially confusing other layers. tcp_ns_to_ts() is changed to return a full 64bit value, ts and ts_now variables are changed to u64 type, and TSMASK is removed in favor of shifts operations. While we are at it, change this sequence: ts >>= TSBITS; ts--; ts <<= TSBITS; ts |= options; to: ts -= (1UL << TSBITS); Fixes: 9a568de4818d ("tcp: switch TCP TS option (RFC 7323) to 1ms clock") Signed-off-by: Eric Dumazet Signed-off-by: David S. Miller Signed-off-by: Sasha Levin

tcp: fix excessive TLP and RACK timeouts from HZ rounding

2023-10-25T10:16:14+00:00

commit 1c2709cfff1dedbb9591e989e2f001484208d914 upstream. We discovered from packet traces of slow loss recovery on kernels with the default HZ=250 setting (and min_rtt < 1ms) that after reordering, when receiving a SACKed sequence range, the RACK reordering timer was firing after about 16ms rather than the desired value of roughly min_rtt/4 + 2ms. The problem is largely due to the RACK reorder timer calculation adding in TCP_TIMEOUT_MIN, which is 2 jiffies. On kernels with HZ=250, this is 2*4ms = 8ms. The TLP timer calculation has the exact same issue. This commit fixes the TLP transmit timer and RACK reordering timer floor calculation to more closely match the intended 2ms floor even on kernels with HZ=250. It does this by adding in a new TCP_TIMEOUT_MIN_US floor of 2000 us and then converting to jiffies, instead of the current approach of converting to jiffies and then adding th TCP_TIMEOUT_MIN value of 2 jiffies. Our testing has verified that on kernels with HZ=1000, as expected, this does not produce significant changes in behavior, but on kernels with the default HZ=250 the latency improvement can be large. For example, our tests show that for HZ=250 kernels at low RTTs this fix roughly halves the latency for the RACK reorder timer: instead of mostly firing at 16ms it mostly fires at 8ms. Suggested-by: Eric Dumazet Signed-off-by: Neal Cardwell Signed-off-by: Yuchung Cheng Fixes: bb4d991a28cc ("tcp: adjust tail loss probe timeout") Reviewed-by: Eric Dumazet Link: https://lore.kernel.org/r/20231015174700.2206872-1-ncardwell.sw@gmail.com Signed-off-by: Jakub Kicinski Signed-off-by: Greg Kroah-Hartman

tcp: fix quick-ack counting to count actual ACKs of new data

2023-10-10T20:03:03+00:00

[ Upstream commit 059217c18be6757b95bfd77ba53fb50b48b8a816 ] This commit fixes quick-ack counting so that it only considers that a quick-ack has been provided if we are sending an ACK that newly acknowledges data. The code was erroneously using the number of data segments in outgoing skbs when deciding how many quick-ack credits to remove. This logic does not make sense, and could cause poor performance in request-response workloads, like RPC traffic, where requests or responses can be multi-segment skbs. When a TCP connection decides to send N quick-acks, that is to accelerate the cwnd growth of the congestion control module controlling the remote endpoint of the TCP connection. That quick-ack decision is purely about the incoming data and outgoing ACKs. It has nothing to do with the outgoing data or the size of outgoing data. And in particular, an ACK only serves the intended purpose of allowing the remote congestion control to grow the congestion window quickly if the ACK is ACKing or SACKing new data. The fix is simple: only count packets as serving the goal of the quickack mechanism if they are ACKing/SACKing new data. We can tell whether this is the case by checking inet_csk_ack_scheduled(), since we schedule an ACK exactly when we are ACKing/SACKing new data. Fixes: fc6415bcb0f5 ("[TCP]: Fix quick-ack decrementing with TSO.") Signed-off-by: Neal Cardwell Reviewed-by: Yuchung Cheng Reviewed-by: Eric Dumazet Link: https://lore.kernel.org/r/20231001151239.1866845-1-ncardwell.sw@gmail.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

tcp: tcp_enter_quickack_mode() should be static

2023-09-13T07:53:02+00:00

[ Upstream commit 03b123debcbc8db987bda17ed8412cc011064c22 ] After commit d2ccd7bc8acd ("tcp: avoid resetting ACK timer in DCTCP"), tcp_enter_quickack_mode() is only used from net/ipv4/tcp_input.c. Fixes: d2ccd7bc8acd ("tcp: avoid resetting ACK timer in DCTCP") Signed-off-by: Eric Dumazet Cc: Yuchung Cheng Cc: Neal Cardwell Link: https://lore.kernel.org/r/20230718162049.1444938-1-edumazet@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

tcp: annotate data-races around tp->notsent_lowat

2023-07-20T19:34:18+00:00

tp->notsent_lowat can be read locklessly from do_tcp_getsockopt() and tcp_poll(). Fixes: c9bee3b7fdec ("tcp: TCP_NOTSENT_LOWAT socket option") Signed-off-by: Eric Dumazet Link: https://lore.kernel.org/r/20230719212857.3943972-10-edumazet@google.com Signed-off-by: Jakub Kicinski

tcp: annotate data-races around tp->keepalive_probes

2023-07-20T19:34:18+00:00

do_tcp_getsockopt() reads tp->keepalive_probes while another cpu might change its value. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Signed-off-by: Eric Dumazet Link: https://lore.kernel.org/r/20230719212857.3943972-6-edumazet@google.com Signed-off-by: Jakub Kicinski

tcp: annotate data-races around tp->keepalive_intvl

2023-07-20T19:34:18+00:00

do_tcp_getsockopt() reads tp->keepalive_intvl while another cpu might change its value. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Signed-off-by: Eric Dumazet Link: https://lore.kernel.org/r/20230719212857.3943972-5-edumazet@google.com Signed-off-by: Jakub Kicinski

tcp: annotate data-races around tp->keepalive_time

2023-07-20T19:34:18+00:00

do_tcp_getsockopt() reads tp->keepalive_time while another cpu might change its value. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Signed-off-by: Eric Dumazet Link: https://lore.kernel.org/r/20230719212857.3943972-4-edumazet@google.com Signed-off-by: Jakub Kicinski

sock: Remove ->sendpage*() in favour of sendmsg(MSG_SPLICE_PAGES)

2023-06-24T22:50:13+00:00

Remove ->sendpage() and ->sendpage_locked(). sendmsg() with MSG_SPLICE_PAGES should be used instead. This allows multiple pages and multipage folios to be passed through. Signed-off-by: David Howells Acked-by: Marc Kleine-Budde # for net/can cc: Jens Axboe cc: Matthew Wilcox cc: linux-afs@lists.infradead.org cc: mptcp@lists.linux.dev cc: rds-devel@oss.oracle.com cc: tipc-discussion@lists.sourceforge.net cc: virtualization@lists.linux-foundation.org Link: https://lore.kernel.org/r/20230623225513.2732256-16-dhowells@redhat.com Signed-off-by: Jakub Kicinski

tcp: Use per-vma locking for receive zerocopy

2023-06-18T10:16:00+00:00

Per-VMA locking allows us to lock a struct vm_area_struct without taking the process-wide mmap lock in read mode. Consider a process workload where the mmap lock is taken constantly in write mode. In this scenario, all zerocopy receives are periodically blocked during that period of time - though in principle, the memory ranges being used by TCP are not touched by the operations that need the mmap write lock. This results in performance degradation. Now consider another workload where the mmap lock is never taken in write mode, but there are many TCP connections using receive zerocopy that are concurrently receiving. These connections all take the mmap lock in read mode, but this does induce a lot of contention and atomic ops for this process-wide lock. This results in additional CPU overhead caused by contending on the cache line for this lock. However, with per-vma locking, both of these problems can be avoided. As a test, I ran an RPC-style request/response workload with 4KB payloads and receive zerocopy enabled, with 100 simultaneous TCP connections. I measured perf cycles within the find_tcp_vma/mmap_read_lock/mmap_read_unlock codepath, with and without per-vma locking enabled. When using process-wide mmap semaphore read locking, about 1% of measured perf cycles were within this path. With per-VMA locking, this value dropped to about 0.45%. Signed-off-by: Arjun Roy Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller