| Message ID | 20210809185314.38187-1-tom@herbertland.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | txhash: Make hash rethink configurable and change the default | expand |
On Mon, Aug 9, 2021 at 11:53 AM Tom Herbert <tom@herbertland.com> wrote: > > Alexander Azimov performed some nice analysis of the feature in Linux > stack where the IPv6 flow label is changed when the stack detects a > connection is failing. The idea of the algorithm is to try to find a > better path. His reults are quite impressive, and show that this form > of source routing can work effectively. > > Alex raised an issue in that if the server endpoint is an IP anycast > address, the connection might break if the flow label changes routing > of packets on the connection. Anycast is known to be susceptible to > route changes, not just those caused be flow label. The concern is that > flow label modulation might increases the chances that anycast > connections might break, especially if the rethink occurs after just > one RTO which is the current behavior. > > This patch set makes the rethink behavior granular and configurable. > It allows control of when to do the hash rethink: upon negative advice, > at RTO in SYN state, at RTO when not in SYN state. The behavior can > be configured by sysctl and by a socket option. > > This patch set the defautl rethink behavior to be to do a rethink only > on negative advice. This is reverts back to the original behavior of > the hash rethink mechanism. This less aggressive with the intent of Thanks for offering knobs to the txhash mechanism. Any reason why reverting the default behavior (that was changed in 2013) is necessary? systems now rely on this RTO tx-rehash to work around link failures will now have to manually re-enable it. Some users may have to learn from higher connection failures to eventually identify this kernel change. > mitigating potentail breakages when anycast addresses are present.> For those users that are benefitting from changing the hash at the > first RTO, they would retain that behavior by setting the sysctl. > *** BLURB HERE *** > > Tom Herbert (3): > txhash: Make rethinking txhash behavior configurable via sysctl > txhash: Add socket option to control TX hash rethink behavior > txhash: Change default rethink behavior to be less aggressive > > arch/alpha/include/uapi/asm/socket.h | 2 ++ > arch/mips/include/uapi/asm/socket.h | 2 ++ > arch/parisc/include/uapi/asm/socket.h | 2 ++ > arch/sparc/include/uapi/asm/socket.h | 3 ++- > include/net/netns/core.h | 2 ++ > include/net/sock.h | 32 +++++++++++++++++++-------- > include/uapi/asm-generic/socket.h | 2 ++ > include/uapi/linux/socket.h | 13 +++++++++++ > net/core/net_namespace.c | 4 ++++ > net/core/sock.c | 16 ++++++++++++++ > net/core/sysctl_net_core.c | 7 ++++++ > net/ipv4/tcp_input.c | 2 +- > net/ipv4/tcp_timer.c | 5 ++++- > 13 files changed, 80 insertions(+), 12 deletions(-) > > -- > 2.25.1 >
On Mon, Aug 9, 2021 at 2:56 PM Yuchung Cheng <ycheng@google.com> wrote: > > On Mon, Aug 9, 2021 at 11:53 AM Tom Herbert <tom@herbertland.com> wrote: > > > > Alexander Azimov performed some nice analysis of the feature in Linux > > stack where the IPv6 flow label is changed when the stack detects a > > connection is failing. The idea of the algorithm is to try to find a > > better path. His reults are quite impressive, and show that this form > > of source routing can work effectively. > > > > Alex raised an issue in that if the server endpoint is an IP anycast > > address, the connection might break if the flow label changes routing > > of packets on the connection. Anycast is known to be susceptible to > > route changes, not just those caused be flow label. The concern is that > > flow label modulation might increases the chances that anycast > > connections might break, especially if the rethink occurs after just > > one RTO which is the current behavior. > > > > This patch set makes the rethink behavior granular and configurable. > > It allows control of when to do the hash rethink: upon negative advice, > > at RTO in SYN state, at RTO when not in SYN state. The behavior can > > be configured by sysctl and by a socket option. > > > > This patch set the defautl rethink behavior to be to do a rethink only > > on negative advice. This is reverts back to the original behavior of > > the hash rethink mechanism. This less aggressive with the intent of > Thanks for offering knobs to the txhash mechanism. > > Any reason why reverting the default behavior (that was changed in > 2013) is necessary? systems now rely on this RTO tx-rehash to work > around link failures will now have to manually re-enable it. Some > users may have to learn from higher connection failures to eventually > identify this kernel change. Just to be clear: I agree we should offer knobs to change the txhash behavior so the first parts of this set looks good to me. I am only concerned about the default behavior reversal. > > > mitigating potentail breakages when anycast addresses are present.> For those users that are benefitting from changing the hash at the > > first RTO, they would retain that behavior by setting the sysctl. > > *** BLURB HERE *** > > > > Tom Herbert (3): > > txhash: Make rethinking txhash behavior configurable via sysctl > > txhash: Add socket option to control TX hash rethink behavior > > txhash: Change default rethink behavior to be less aggressive > > > > arch/alpha/include/uapi/asm/socket.h | 2 ++ > > arch/mips/include/uapi/asm/socket.h | 2 ++ > > arch/parisc/include/uapi/asm/socket.h | 2 ++ > > arch/sparc/include/uapi/asm/socket.h | 3 ++- > > include/net/netns/core.h | 2 ++ > > include/net/sock.h | 32 +++++++++++++++++++-------- > > include/uapi/asm-generic/socket.h | 2 ++ > > include/uapi/linux/socket.h | 13 +++++++++++ > > net/core/net_namespace.c | 4 ++++ > > net/core/sock.c | 16 ++++++++++++++ > > net/core/sysctl_net_core.c | 7 ++++++ > > net/ipv4/tcp_input.c | 2 +- > > net/ipv4/tcp_timer.c | 5 ++++- > > 13 files changed, 80 insertions(+), 12 deletions(-) > > > > -- > > 2.25.1 > >