Message ID | 20230511093456.672221-5-atenart@kernel.org (mailing list archive) |
---|---|
State | Superseded |
Delegated to: | Netdev Maintainers |
Headers | show |
Series | net: tcp: make txhash use consistent for IPv4 | expand |
Hi Antoine, On 5/11/23 11:34, Antoine Tenart wrote: > Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") > sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is > used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. > With this, skb->l4_hash does not always indicate the hash is a > "canonical 4-tuple hash over transport ports" but rather a hash from L4 > layer to provide a uniform distribution over flows. Reword the comment > accordingly, to avoid misunderstandings. But AFAIU the hash used to be a canonical 4-tuple hash and was used as such by other components, e.g., OvS: https://elixir.bootlin.com/linux/latest/source/net/openvswitch/actions.c#L1069 It seems to me at least unfortunate that semantics change without considering other users. The fact that we now fix the documentation makes it seem like OvS was wrong to use the skb hash. However, before 877d1f6291f8 ("net: Set sk_txhash from a random number") it was OK for OvS to use the skb hash as a canonical 4-tuple hash. Best regards, Dumitru > > Signed-off-by: Antoine Tenart <atenart@kernel.org> > --- > include/linux/skbuff.h | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h > index 738776ab8838..f54c84193b23 100644 > --- a/include/linux/skbuff.h > +++ b/include/linux/skbuff.h > @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; > * @active_extensions: active extensions (skb_ext_id types) > * @ndisc_nodetype: router type (from link layer) > * @ooo_okay: allow the mapping of a socket to a queue to be changed > - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport > - * ports. > + * @l4_hash: indicate hash is from layer 4 and provides a uniform > + * distribution over flows. > * @sw_hash: indicates hash was computed in software stack > * @wifi_acked_valid: wifi_acked was set > * @wifi_acked: whether frame was acked on wifi or not
On Thu, May 11, 2023 at 2:10 PM Dumitru Ceara <dceara@redhat.com> wrote: > > Hi Antoine, > > On 5/11/23 11:34, Antoine Tenart wrote: > > Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") > > sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is > > used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. > > With this, skb->l4_hash does not always indicate the hash is a > > "canonical 4-tuple hash over transport ports" but rather a hash from L4 > > layer to provide a uniform distribution over flows. Reword the comment > > accordingly, to avoid misunderstandings. > > But AFAIU the hash used to be a canonical 4-tuple hash and was used as > such by other components, e.g., OvS: > > https://elixir.bootlin.com/linux/latest/source/net/openvswitch/actions.c#L1069 > > It seems to me at least unfortunate that semantics change without > considering other users. The fact that we now fix the documentation > makes it seem like OvS was wrong to use the skb hash. However, before > 877d1f6291f8 ("net: Set sk_txhash from a random number") it was OK for > OvS to use the skb hash as a canonical 4-tuple hash. > I do not think we can undo stuff that was done back in 2015 Has anyone complained ? Note that skb->hash has never been considered as canonical, for obvious reasons. > Best regards, > Dumitru > > > > > Signed-off-by: Antoine Tenart <atenart@kernel.org> > > --- > > include/linux/skbuff.h | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h > > index 738776ab8838..f54c84193b23 100644 > > --- a/include/linux/skbuff.h > > +++ b/include/linux/skbuff.h > > @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; > > * @active_extensions: active extensions (skb_ext_id types) > > * @ndisc_nodetype: router type (from link layer) > > * @ooo_okay: allow the mapping of a socket to a queue to be changed > > - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport > > - * ports. > > + * @l4_hash: indicate hash is from layer 4 and provides a uniform > > + * distribution over flows. > > * @sw_hash: indicates hash was computed in software stack > > * @wifi_acked_valid: wifi_acked was set > > * @wifi_acked: whether frame was acked on wifi or not >
On 5/11/23 14:33, Eric Dumazet wrote: > On Thu, May 11, 2023 at 2:10 PM Dumitru Ceara <dceara@redhat.com> wrote: >> >> Hi Antoine, >> >> On 5/11/23 11:34, Antoine Tenart wrote: >>> Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") >>> sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is >>> used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. >>> With this, skb->l4_hash does not always indicate the hash is a >>> "canonical 4-tuple hash over transport ports" but rather a hash from L4 >>> layer to provide a uniform distribution over flows. Reword the comment >>> accordingly, to avoid misunderstandings. >> >> But AFAIU the hash used to be a canonical 4-tuple hash and was used as >> such by other components, e.g., OvS: >> >> https://elixir.bootlin.com/linux/latest/source/net/openvswitch/actions.c#L1069 >> >> It seems to me at least unfortunate that semantics change without >> considering other users. The fact that we now fix the documentation >> makes it seem like OvS was wrong to use the skb hash. However, before >> 877d1f6291f8 ("net: Set sk_txhash from a random number") it was OK for >> OvS to use the skb hash as a canonical 4-tuple hash. >> > > I do not think we can undo stuff that was done back in 2015 > I understand. I guess I was kind of grasping at straws in the hope of getting a canonical 4-tuple hash. > Has anyone complained ? > It did go unnoticed for a while but recently we started getting (indirect) reports due to the hash changing. This one is from an upstream OVN (OvS) user: https://github.com/ovn-org/ovn/issues/112 This is from an OpenShift (also running OVN/OvS) user: https://issues.redhat.com/browse/OCPBUGS-7406 > Note that skb->hash has never been considered as canonical, for obvious reasons. > > >> Best regards, >> Dumitru >> >>> >>> Signed-off-by: Antoine Tenart <atenart@kernel.org> >>> --- >>> include/linux/skbuff.h | 4 ++-- >>> 1 file changed, 2 insertions(+), 2 deletions(-) >>> >>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h >>> index 738776ab8838..f54c84193b23 100644 >>> --- a/include/linux/skbuff.h >>> +++ b/include/linux/skbuff.h >>> @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; >>> * @active_extensions: active extensions (skb_ext_id types) >>> * @ndisc_nodetype: router type (from link layer) >>> * @ooo_okay: allow the mapping of a socket to a queue to be changed >>> - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport >>> - * ports. >>> + * @l4_hash: indicate hash is from layer 4 and provides a uniform >>> + * distribution over flows. >>> * @sw_hash: indicates hash was computed in software stack >>> * @wifi_acked_valid: wifi_acked was set >>> * @wifi_acked: whether frame was acked on wifi or not >> >
On 5/11/23 15:00, Dumitru Ceara wrote: > On 5/11/23 14:33, Eric Dumazet wrote: >> On Thu, May 11, 2023 at 2:10 PM Dumitru Ceara <dceara@redhat.com> wrote: >>> >>> Hi Antoine, >>> >>> On 5/11/23 11:34, Antoine Tenart wrote: >>>> Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") >>>> sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is >>>> used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. >>>> With this, skb->l4_hash does not always indicate the hash is a >>>> "canonical 4-tuple hash over transport ports" but rather a hash from L4 >>>> layer to provide a uniform distribution over flows. Reword the comment >>>> accordingly, to avoid misunderstandings. >>> >>> But AFAIU the hash used to be a canonical 4-tuple hash and was used as >>> such by other components, e.g., OvS: >>> >>> https://elixir.bootlin.com/linux/latest/source/net/openvswitch/actions.c#L1069 >>> >>> It seems to me at least unfortunate that semantics change without >>> considering other users. The fact that we now fix the documentation >>> makes it seem like OvS was wrong to use the skb hash. However, before >>> 877d1f6291f8 ("net: Set sk_txhash from a random number") it was OK for >>> OvS to use the skb hash as a canonical 4-tuple hash. >>> >> >> I do not think we can undo stuff that was done back in 2015 >> > > I understand. I guess I was kind of grasping at straws in the hope of > getting a canonical 4-tuple hash. > >> Has anyone complained ? >> > > It did go unnoticed for a while but recently we started getting > (indirect) reports due to the hash changing. > > This one is from an upstream OVN (OvS) user: > https://github.com/ovn-org/ovn/issues/112 > > This is from an OpenShift (also running OVN/OvS) user: > https://issues.redhat.com/browse/OCPBUGS-7406 > >> Note that skb->hash has never been considered as canonical, for obvious reasons. I guess, the other point here is that it's not an L4 hash either. It's a random number. So, the documentation will still not be correct even after the change proposed in this patch. One other solution to the problem might be to stop setting l4_hash flag while it's a random number. One way to not break everything doing that will be to introduce a new flag, e.g. 'rnd_hash' that will be a hash that is "not related to packet fields, but provides a uniform distribution over flows". skb_get_hash() then may return the current hash if it's any of l4, rnd or sw. That should preserve the current logic across the kernel code. But having a new flag, we could introduce a new helper, for example skb_get_stable_hash() or skb_get_hash_nonrandom() or something like that, that will be equal to the current version of skb_get_hash(), i.e. not take the random hash into account. Affected subsystems (OVS, ECMP, SRv6) can be changed to use that new function. This way these subsystems will get a software hash based on the real packet fields, if it was originally random. This will also preserve ability to use hash provided by the HW, since it is not normally random. With that, we'll also not need to have in the API something that has 'L4' in the name and in the docs, but has no relation to packet fields. It can be argued that the description in the doc doesn't mean that this hash is computed using L4 packet fields, but it's confusing regardless and getting overlooked while creating new code, as it shown by the issues in multiple substystems. Hope this makes some sense. Dumitru also had some alternative ideas on how to provide a stable hash to subsystems that need it, but I'll leave it to him. Best regards, Ilya Maximets. >> >> >>> Best regards, >>> Dumitru >>> >>>> >>>> Signed-off-by: Antoine Tenart <atenart@kernel.org> >>>> --- >>>> include/linux/skbuff.h | 4 ++-- >>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>> >>>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h >>>> index 738776ab8838..f54c84193b23 100644 >>>> --- a/include/linux/skbuff.h >>>> +++ b/include/linux/skbuff.h >>>> @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; >>>> * @active_extensions: active extensions (skb_ext_id types) >>>> * @ndisc_nodetype: router type (from link layer) >>>> * @ooo_okay: allow the mapping of a socket to a queue to be changed >>>> - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport >>>> - * ports. >>>> + * @l4_hash: indicate hash is from layer 4 and provides a uniform >>>> + * distribution over flows. >>>> * @sw_hash: indicates hash was computed in software stack >>>> * @wifi_acked_valid: wifi_acked was set >>>> * @wifi_acked: whether frame was acked on wifi or not >>> >> >
On 5/11/23 19:54, Ilya Maximets wrote: > On 5/11/23 15:00, Dumitru Ceara wrote: >> On 5/11/23 14:33, Eric Dumazet wrote: >>> On Thu, May 11, 2023 at 2:10 PM Dumitru Ceara <dceara@redhat.com> wrote: >>>> >>>> Hi Antoine, >>>> >>>> On 5/11/23 11:34, Antoine Tenart wrote: >>>>> Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") >>>>> sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is >>>>> used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. >>>>> With this, skb->l4_hash does not always indicate the hash is a >>>>> "canonical 4-tuple hash over transport ports" but rather a hash from L4 >>>>> layer to provide a uniform distribution over flows. Reword the comment >>>>> accordingly, to avoid misunderstandings. >>>> >>>> But AFAIU the hash used to be a canonical 4-tuple hash and was used as >>>> such by other components, e.g., OvS: >>>> >>>> https://elixir.bootlin.com/linux/latest/source/net/openvswitch/actions.c#L1069 >>>> >>>> It seems to me at least unfortunate that semantics change without >>>> considering other users. The fact that we now fix the documentation >>>> makes it seem like OvS was wrong to use the skb hash. However, before >>>> 877d1f6291f8 ("net: Set sk_txhash from a random number") it was OK for >>>> OvS to use the skb hash as a canonical 4-tuple hash. >>>> >>> >>> I do not think we can undo stuff that was done back in 2015 >>> >> >> I understand. I guess I was kind of grasping at straws in the hope of >> getting a canonical 4-tuple hash. >> >>> Has anyone complained ? >>> >> >> It did go unnoticed for a while but recently we started getting >> (indirect) reports due to the hash changing. >> >> This one is from an upstream OVN (OvS) user: >> https://github.com/ovn-org/ovn/issues/112 >> >> This is from an OpenShift (also running OVN/OvS) user: >> https://issues.redhat.com/browse/OCPBUGS-7406 >> I just realized we need a bit more context here. It started being a visible problem after 265f94ff54d6 ("net: Recompute sk_txhash on negative routing advice") and also after 3acf3ec3f4b0 ("tcp: Change txhash on every SYN and RTO retransmit") when retransmits started changing the txhash and implicitly the hash used by OvS. >>> Note that skb->hash has never been considered as canonical, for obvious reasons. > > I guess, the other point here is that it's not an L4 hash either. > > It's a random number. So, the documentation will still not be > correct even after the change proposed in this patch. > > > One other solution to the problem might be to stop setting l4_hash > flag while it's a random number. > > One way to not break everything doing that will be to introduce a > new flag, e.g. 'rnd_hash' that will be a hash that is "not related > to packet fields, but provides a uniform distribution over flows". > > skb_get_hash() then may return the current hash if it's any of > l4, rnd or sw. That should preserve the current logic across > the kernel code. > But having a new flag, we could introduce a new helper, for example > skb_get_stable_hash() or skb_get_hash_nonrandom() or something like > that, that will be equal to the current version of skb_get_hash(), > i.e. not take the random hash into account. > > Affected subsystems (OVS, ECMP, SRv6) can be changed to use that > new function. This way these subsystems will get a software hash > based on the real packet fields, if it was originally random. > This will also preserve ability to use hash provided by the HW, > since it is not normally random. > > With that, we'll also not need to have in the API something that has > 'L4' in the name and in the docs, but has no relation to packet fields. > It can be argued that the description in the doc doesn't mean that > this hash is computed using L4 packet fields, but it's confusing > regardless and getting overlooked while creating new code, as it > shown by the issues in multiple substystems. > > Hope this makes some sense. > > > Dumitru also had some alternative ideas on how to provide a stable > hash to subsystems that need it, but I'll leave it to him. > What I had in mind is not really a stable hash but a "good enough alternative". It's probably "good enough" (at least for OvS/OVN) if the hash used by OvS doesn't change throughout the lifetime of a TCP session. Would it be possible to save the original (random) hash that was generated for a locally terminated TCP session? E.g., a new field in 'struct sock'. It would be in essence a random tag associated to the session that doesn't change throughout the lifetime of the session. Unlike sk->sk_txhash which changes on retransmit/negative routing advice. That means OvS doesn't have to compute a stable hash every time it processes a packet, It would just access this value through skb->sk->good_name_for_this_new_tag. The advantage is that it gives the appearance of a canonical 4-tuple hash throughout the lifetime of a session and it doesn't affect any of the use cases that required 877d1f6291f8 ("net: Set sk_txhash from a random number"). I probably missed relevant things but I thought it might be worth sharing in case the idea has some value. Regards, Dumitru > Best regards, Ilya Maximets. > >>> >>> >>>> Best regards, >>>> Dumitru >>>> >>>>> >>>>> Signed-off-by: Antoine Tenart <atenart@kernel.org> >>>>> --- >>>>> include/linux/skbuff.h | 4 ++-- >>>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>>> >>>>> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h >>>>> index 738776ab8838..f54c84193b23 100644 >>>>> --- a/include/linux/skbuff.h >>>>> +++ b/include/linux/skbuff.h >>>>> @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; >>>>> * @active_extensions: active extensions (skb_ext_id types) >>>>> * @ndisc_nodetype: router type (from link layer) >>>>> * @ooo_okay: allow the mapping of a socket to a queue to be changed >>>>> - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport >>>>> - * ports. >>>>> + * @l4_hash: indicate hash is from layer 4 and provides a uniform >>>>> + * distribution over flows. >>>>> * @sw_hash: indicates hash was computed in software stack >>>>> * @wifi_acked_valid: wifi_acked was set >>>>> * @wifi_acked: whether frame was acked on wifi or not >>>> >>> >> >
Quoting Dumitru Ceara (2023-05-11 22:50:32) > On 5/11/23 19:54, Ilya Maximets wrote: > >>> Note that skb->hash has never been considered as canonical, for obvious reasons. > > > > I guess, the other point here is that it's not an L4 hash either. > > > > It's a random number. So, the documentation will still not be > > correct even after the change proposed in this patch. The proposed changed is "indicate hash is from layer 4 and provides a uniform distribution over flows", which does not describe *how* the hash is computed but *where* it comes from. This matches "random number set by TCP" and changes in how hashes are computed won't affect the comment, so we'll not end up in the same situation. > > One way to not break everything doing that will be to introduce a > > new flag, e.g. 'rnd_hash' that will be a hash that is "not related > > to packet fields, but provides a uniform distribution over flows". > > > > skb_get_hash() then may return the current hash if it's any of > > l4, rnd or sw. That should preserve the current logic across > > the kernel code. > > But having a new flag, we could introduce a new helper, for example > > skb_get_stable_hash() or skb_get_hash_nonrandom() or something like > > that, that will be equal to the current version of skb_get_hash(), > > i.e. not take the random hash into account. > > > > Affected subsystems (OVS, ECMP, SRv6) can be changed to use that > > new function. This way these subsystems will get a software hash > > based on the real packet fields, if it was originally random. > > This will also preserve ability to use hash provided by the HW, > > since it is not normally random. But then the whole point of txrehash would be dismissed, if ECMP and others stop using the hash provided by TCP. This needs to be a conditional setting, to make the skb hash to be stable over time only when needed. That way both scenario are supported. > What I had in mind is not really a stable hash but a "good enough > alternative". It's probably "good enough" (at least for OvS/OVN) if the > hash used by OvS doesn't change throughout the lifetime of a TCP session. So what's important is not how the hash is computed but the fact that it should be stable over time when requested. Isn't exactly what net.core.txrehash=0 does? If there are some bugs they should be fixed. On top of this, if OvS needs to additionally provide a canonical 4/5-tuple hash because not only the stability over time is needed but also the method is important, it needs to compute its own hash. As part of such potential series ways to cache the result can be explored. Numbers would help too. (This can be discussed here, that's fine, but I thought it's important to distinguish the two topics). Antoine
On 5/15/23 10:12, Antoine Tenart wrote: > Quoting Dumitru Ceara (2023-05-11 22:50:32) >> On 5/11/23 19:54, Ilya Maximets wrote: >>>>> Note that skb->hash has never been considered as canonical, for obvious reasons. >>> >>> I guess, the other point here is that it's not an L4 hash either. >>> >>> It's a random number. So, the documentation will still not be >>> correct even after the change proposed in this patch. > > The proposed changed is "indicate hash is from layer 4 and provides a > uniform distribution over flows", which does not describe *how* the hash > is computed but *where* it comes from. This matches "random number set > by TCP" and changes in how hashes are computed won't affect the comment, > so we'll not end up in the same situation. I respectfully disagree, "is from layer 4" and "random number" do not match for me. So, "where it comes from" argument is not applicable. Random numbers come from random number generator, and not "from layer 4". Unless by "from layer 4" you mean "from the code that handles layer 4 packet processing". But that seems very confusing to me. And it is definitely not the first thing that comes to mind while reading the documentation. > >>> One way to not break everything doing that will be to introduce a >>> new flag, e.g. 'rnd_hash' that will be a hash that is "not related >>> to packet fields, but provides a uniform distribution over flows". >>> >>> skb_get_hash() then may return the current hash if it's any of >>> l4, rnd or sw. That should preserve the current logic across >>> the kernel code. >>> But having a new flag, we could introduce a new helper, for example >>> skb_get_stable_hash() or skb_get_hash_nonrandom() or something like >>> that, that will be equal to the current version of skb_get_hash(), >>> i.e. not take the random hash into account. >>> >>> Affected subsystems (OVS, ECMP, SRv6) can be changed to use that >>> new function. This way these subsystems will get a software hash >>> based on the real packet fields, if it was originally random. >>> This will also preserve ability to use hash provided by the HW, >>> since it is not normally random. > > But then the whole point of txrehash would be dismissed, if ECMP and > others stop using the hash provided by TCP. I guess ECMP is not a good example as it doesn't use skb hash for locally generated traffic. However, the argument about defeating the purpose of rehash doesn't stand either for the same reason. The same next hop will be chosen for the same flow always, because the hash will be re-generated from a 5-tuple. But anyway... In OVS and SRv6 cases we're also talking about load balancing, so unlike ECMP, final destinations may be different based on the hash. In this case there is an issue. > This needs to be a > conditional setting, to make the skb hash to be stable over time only > when needed. That way both scenario are supported. > >> What I had in mind is not really a stable hash but a "good enough >> alternative". It's probably "good enough" (at least for OvS/OVN) if the >> hash used by OvS doesn't change throughout the lifetime of a TCP session. > > So what's important is not how the hash is computed but the fact that > it should be stable over time when requested. Isn't exactly what > net.core.txrehash=0 does? If there are some bugs they should be fixed. Not asking to disable re-hash. Asking to provide a way to distinguish changing random numbers from a more stable hash, so we can avoid recalculating it on every packet. Of course, we can change the code to re-calculate every time, but that sounds wasteful if it's already done by the HW, for example. > On top of this, if OvS needs to additionally provide a canonical > 4/5-tuple hash because not only the stability over time is needed but > also the method is important, it needs to compute its own hash For the OVS_HASH_ALG_L4, the exact method is not really important. For OVS_HASH_ALG_SYM_L4, the hash has to symmetric, but this hashing algorithm is not implemented in the OVS kernel module today. > As part > of such potential series ways to cache the result can be explored. > Numbers would help too. (This can be discussed here, that's fine, but I > thought it's important to distinguish the two topics). I agree that topics are fairly different. The rest of the set is probably fine (I didn't review). I'm just not sure why we need to change the comment from one incorrect sentence to another similarly incorrect and confusing. Also making it look like subsystems that use it in a previously documented meaning are at fault. It's just not fair. I think, the issue should be fixed first. On the other note, I don't think that the rest of the patch set should be held back by that though. So, maybe dropping this one patch from the set might be an option for now? Best regards, Ilya Maximets.
Quoting Ilya Maximets (2023-05-15 20:23:28) > On 5/15/23 10:12, Antoine Tenart wrote: > > Quoting Dumitru Ceara (2023-05-11 22:50:32) > >> On 5/11/23 19:54, Ilya Maximets wrote: > >>>>> Note that skb->hash has never been considered as canonical, for obvious reasons. > >>> > >>> I guess, the other point here is that it's not an L4 hash either. > >>> > >>> It's a random number. So, the documentation will still not be > >>> correct even after the change proposed in this patch. > > > > The proposed changed is "indicate hash is from layer 4 and provides a > > uniform distribution over flows", which does not describe *how* the hash > > is computed but *where* it comes from. This matches "random number set > > by TCP" and changes in how hashes are computed won't affect the comment, > > so we'll not end up in the same situation. > > I respectfully disagree, "is from layer 4" and "random number" do not > match for me. So, "where it comes from" argument is not applicable. > Random numbers come from random number generator, and not "from layer 4". > > Unless by "from layer 4" you mean "from the code that handles layer 4 > packet processing". But that seems very confusing to me. And it is > definitely not the first thing that comes to mind while reading the > documentation. Yes that is what I meant, but if that is still confusing then this is not improving things so let's try something better. I intentionally did not mention how the hash is computed because it's easy to forget to update the documentation when the exact logic is changed. What's important here IMHO is to mention what the hash provides. What about "indicates hash was set by layer 4 stack and provides a uniform distribution over flows"? Or/and we should we also add a disclaimer like "no guarantee on how the hash was computed"? > Also making it look like subsystems that use it in a previously > documented meaning are at fault. It's just not fair. I never said that nor it is what I think, I'm sorry if my message was misleading. Antoine
On 5/16/23 09:36, Antoine Tenart wrote: > Quoting Ilya Maximets (2023-05-15 20:23:28) >> On 5/15/23 10:12, Antoine Tenart wrote: >>> Quoting Dumitru Ceara (2023-05-11 22:50:32) >>>> On 5/11/23 19:54, Ilya Maximets wrote: >>>>>>> Note that skb->hash has never been considered as canonical, for obvious reasons. >>>>> >>>>> I guess, the other point here is that it's not an L4 hash either. >>>>> >>>>> It's a random number. So, the documentation will still not be >>>>> correct even after the change proposed in this patch. >>> >>> The proposed changed is "indicate hash is from layer 4 and provides a >>> uniform distribution over flows", which does not describe *how* the hash >>> is computed but *where* it comes from. This matches "random number set >>> by TCP" and changes in how hashes are computed won't affect the comment, >>> so we'll not end up in the same situation. >> >> I respectfully disagree, "is from layer 4" and "random number" do not >> match for me. So, "where it comes from" argument is not applicable. >> Random numbers come from random number generator, and not "from layer 4". >> >> Unless by "from layer 4" you mean "from the code that handles layer 4 >> packet processing". But that seems very confusing to me. And it is >> definitely not the first thing that comes to mind while reading the >> documentation. > > Yes that is what I meant, but if that is still confusing then this is > not improving things so let's try something better. I intentionally did > not mention how the hash is computed because it's easy to forget to > update the documentation when the exact logic is changed. What's > important here IMHO is to mention what the hash provides. > > What about "indicates hash was set by layer 4 stack and provides a > uniform distribution over flows"? Or/and we should we also add a > disclaimer like "no guarantee on how the hash was computed"? I'm still not sure this is correct. Is a NIC driver part of layer 4 stack? Also it still doesn't make a lot of sense to change to comment without changing the code. And there are lots of other inconsistencies around skb hash. The following is probably the most colorful that I found: TCP code in tcp_make_synack() calls: skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); Where 'txhash' is a random number. But it calls it PKT_HASH_TYPE_L4. Going to the definition we can find this [1]: /* * Packet hash types specify the type of hash in skb_set_hash. * * Hash types refer to the protocol layer addresses which are used to * construct a packet's hash. The hashes are used to differentiate or identify * flows of the protocol layer for the hash type. Hash types are either * layer-2 (L2), layer-3 (L3), or layer-4 (L4). * * Properties of hashes: * * 1) Two packets in different flows have different hash values * 2) Two packets in the same flow should have the same hash value Now this directly contradicts to the hash being not stable and not being computed form actual packet fields. Later in the same file: enum pkt_hash_types { PKT_HASH_TYPE_NONE, /* Undefined type */ PKT_HASH_TYPE_L2, /* Input: src_MAC, dest_MAC */ PKT_HASH_TYPE_L3, /* Input: src_IP, dst_IP */ PKT_HASH_TYPE_L4, /* Input: src_IP, dst_IP, src_port, dst_port */ }; Here we see that PKT_HASH_TYPE_L4 supposed to use particular fields as an input. It's kind of pointless having all that documented if the l4_hash flag is a random number and none of the kernel subsystems are able to use it in a way it is documented. [1] https://elixir.bootlin.com/linux/v6.4-rc1/source/include/linux/skbuff.h#L1419 Best regards, Ilya Maximets.
Quoting Ilya Maximets (2023-05-16 23:25:19) > On 5/16/23 09:36, Antoine Tenart wrote: > > > > What about "indicates hash was set by layer 4 stack and provides a > > uniform distribution over flows"? Or/and we should we also add a > > disclaimer like "no guarantee on how the hash was computed"? > > I'm still not sure this is correct. Is a NIC driver part of layer 4 > stack? Offloading logic with L4 fields for csum, RSS, etc; we can argue it does something at L4. What about this: "Provides a uniform distribution over L4 flows"? I does look better than the previous proposal IMHO. > And there are lots of other inconsistencies around skb hash. The following > is probably the most colorful that I found: > > skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); > * Hash types refer to the protocol layer addresses which are used to > * construct a packet's hash. The hashes are used to differentiate or identify > * flows of the protocol layer for the hash type. Hash types are either > * layer-2 (L2), layer-3 (L3), or layer-4 (L4). > * > * Properties of hashes: > * > * 1) Two packets in different flows have different hash values > * 2) Two packets in the same flow should have the same hash value > enum pkt_hash_types { > PKT_HASH_TYPE_L4, /* Input: src_IP, dst_IP, src_port, dst_port */ > }; > > Here we see that PKT_HASH_TYPE_L4 supposed to use particular fields > as an input. If we strictly follow the above, do all NIC provide a L4 hash using only the above fields (src_IP, dst_IP, src_port, dst_port)? Having a quick look I'm pretty sure no, both 4 and 5-tuple can be used. What is important is at what level the distribution is. So yes strictly speaking the above PKT_HASH_TYPE_L4 use can be a little surprising, but to me it's a shortcut or a missing update. For perfect correctness we could use __skb_set_hash(skb, tcp_rsk(req)->txhash, false, true) FWIW. Even l4_hash w/o taking the rnd case into account does not guarantee a stable hash for the lifetime of a flow; what happens if packets from the same flow are received on two NICs using different keys and/or algs? Being computed from L4 fields does not mean it is stable. If the stable property is needed, the hash has to be computed locally. And then comes the other topic of caching it for reuse and potential sharing across different consumers, sure. Now, I'll let some time to give a chance for others to chime in. Thanks, Antoine
On 5/17/23 14:05, Antoine Tenart wrote: > Quoting Ilya Maximets (2023-05-16 23:25:19) >> On 5/16/23 09:36, Antoine Tenart wrote: >>> >>> What about "indicates hash was set by layer 4 stack and provides a >>> uniform distribution over flows"? Or/and we should we also add a >>> disclaimer like "no guarantee on how the hash was computed"? >> >> I'm still not sure this is correct. Is a NIC driver part of layer 4 >> stack? > > Offloading logic with L4 fields for csum, RSS, etc; we can argue it does > something at L4. What about this: "Provides a uniform distribution over > L4 flows"? I does look better than the previous proposal IMHO. > >> And there are lots of other inconsistencies around skb hash. The following >> is probably the most colorful that I found: >> >> skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); > >> * Hash types refer to the protocol layer addresses which are used to >> * construct a packet's hash. The hashes are used to differentiate or identify >> * flows of the protocol layer for the hash type. Hash types are either >> * layer-2 (L2), layer-3 (L3), or layer-4 (L4). >> * >> * Properties of hashes: >> * >> * 1) Two packets in different flows have different hash values >> * 2) Two packets in the same flow should have the same hash value > >> enum pkt_hash_types { >> PKT_HASH_TYPE_L4, /* Input: src_IP, dst_IP, src_port, dst_port */ >> }; >> >> Here we see that PKT_HASH_TYPE_L4 supposed to use particular fields >> as an input. > > If we strictly follow the above, do all NIC provide a L4 hash using only > the above fields (src_IP, dst_IP, src_port, dst_port)? Having a quick > look I'm pretty sure no, both 4 and 5-tuple can be used. What is > important is at what level the distribution is. I would read the above as 'at least these fields'. In the continuation of the comment above it's, for example, allowed to set L3 when it is, in fact, L4. > > So yes strictly speaking the above PKT_HASH_TYPE_L4 use can be a little > surprising, but to me it's a shortcut or a missing update. For perfect > correctness we could use > __skb_set_hash(skb, tcp_rsk(req)->txhash, false, true) FWIW. Only after the documentation change. > > Even l4_hash w/o taking the rnd case into account does not guarantee a > stable hash for the lifetime of a flow; what happens if packets from the > same flow are received on two NICs using different keys and/or algs? Following the same logic we can't really say that it "provides a uniform distribution over L4 flows" either. The fact that L4 fields were used to calculate the hash, doesn't mean the hash function is any good. So, the same way as stability can't be part of the definition, the uniform distribution can't be as well. And the 'l4' part of the field name looses the meaning completely. So, we will end up with: * @l4_hash: Some number is stored in the 'hash' field. At this point the is no reason to keep the flag at all. In practice though, such setups are unlikley to be common. Surely, some adequate distribution and some stability for hashes is implied. Not the best possible distribution and not the absolute stability, but good enough to rely on in many cases. And that's why the quoted comment defines "Properties of hashes" this way. > Being computed from L4 fields does not mean it is stable. If the stable > property is needed, the hash has to be computed locally. And then comes > the other topic of caching it for reuse and potential sharing across > different consumers, sure. > > Now, I'll let some time to give a chance for others to chime in. Sure. > > Thanks, > Antoine
Quoting Ilya Maximets (2023-05-18 01:00:40) > On 5/17/23 14:05, Antoine Tenart wrote: > > > > Even l4_hash w/o taking the rnd case into account does not guarantee a > > stable hash for the lifetime of a flow; what happens if packets from the > > same flow are received on two NICs using different keys and/or algs? > > Following the same logic we can't really say that it "provides a uniform > distribution over L4 flows" either. The fact that L4 fields were used > to calculate the hash, doesn't mean the hash function is any good. Well drivers need to either trust the h/w in some ways or not use what is provided if it's broken; or we can't be sure of anything. It's not the same as an example where a valid setup can't guarantee a property by design. > > Now, I'll let some time to give a chance for others to chime in. > > Sure. I don't think we'll get more guidance and we failed to come to an agreement so let's keep this as-is for now; I'll send a v2 w/o this documentation change. As for a way forward and the stability need, IMHO the hash needs to be computed where it is used (with potential cache, not reusing skb->hash) but if you feel this should be addressed at this level a patch doing so might at least get others to comment (both ways). Thanks, Antoine
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h index 738776ab8838..f54c84193b23 100644 --- a/include/linux/skbuff.h +++ b/include/linux/skbuff.h @@ -791,8 +791,8 @@ typedef unsigned char *sk_buff_data_t; * @active_extensions: active extensions (skb_ext_id types) * @ndisc_nodetype: router type (from link layer) * @ooo_okay: allow the mapping of a socket to a queue to be changed - * @l4_hash: indicate hash is a canonical 4-tuple hash over transport - * ports. + * @l4_hash: indicate hash is from layer 4 and provides a uniform + * distribution over flows. * @sw_hash: indicates hash was computed in software stack * @wifi_acked_valid: wifi_acked was set * @wifi_acked: whether frame was acked on wifi or not
Since commit 877d1f6291f8 ("net: Set sk_txhash from a random number") sk->sk_txhash is not a canonical 4-tuple hash. sk->sk_txhash is used in the TCP Tx path to populate skb->hash, with skb->l4_hash=1. With this, skb->l4_hash does not always indicate the hash is a "canonical 4-tuple hash over transport ports" but rather a hash from L4 layer to provide a uniform distribution over flows. Reword the comment accordingly, to avoid misunderstandings. Signed-off-by: Antoine Tenart <atenart@kernel.org> --- include/linux/skbuff.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)