| Message ID | e448120735d71f16ca3e1e845730f7fc29e71ea1.1651861213.git.lucien.xin@gmail.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net] Documentation: add description for net.core.gro_normal_batch | expand |
On Fri, 6 May 2022 14:20:13 -0400 Xin Long wrote: > Describe it in admin-guide/sysctl/net.rst like other Network core options. > Users need to know gro_normal_batch for performance tuning. > > Fixes: 323ebb61e32b ("net: use listified RX for handling GRO_NORMAL skbs") > Reported-by: Prijesh <prpatel@redhat.com> Does Prijesh have a last name? :) > Signed-off-by: Xin Long <lucien.xin@gmail.com> > --- > Documentation/admin-guide/sysctl/net.rst | 10 ++++++++++ > 1 file changed, 10 insertions(+) > > diff --git a/Documentation/admin-guide/sysctl/net.rst b/Documentation/admin-guide/sysctl/net.rst > index f86b5e1623c6..d8a8506f31ad 100644 > --- a/Documentation/admin-guide/sysctl/net.rst > +++ b/Documentation/admin-guide/sysctl/net.rst > @@ -374,6 +374,16 @@ option is set to SOCK_TXREHASH_DEFAULT (i. e. not overridden by setsockopt). > If set to 1 (default), hash rethink is performed on listening socket. > If set to 0, hash rethink is not performed. > > +gro_normal_batch > +---------------- > + > +Maximum number of GRO_NORMAL skbs to batch up for list-RX. When GRO decides > +not to coalesce a packet, instead of passing it to the stack immediately, > +place it on a list. That makes it sounds like only packets which were not coalesced go on the list. IIUC everything goes on that list before traveling up the stack, no? > +Pass this list to the stack at flush time or whenever This sentences is in second person, and the previous one was in third person. > +the number of skbs in this list exceeds gro_normal_batch. s/skbs/segments/ > +Default : 8 > + > 2. /proc/sys/net/unix - Parameters for Unix domain sockets > ---------------------------------------------------------- >
On Mon, May 9, 2022 at 9:07 PM Jakub Kicinski <kuba@kernel.org> wrote: > > On Fri, 6 May 2022 14:20:13 -0400 Xin Long wrote: > > Describe it in admin-guide/sysctl/net.rst like other Network core options. > > Users need to know gro_normal_batch for performance tuning. > > > > Fixes: 323ebb61e32b ("net: use listified RX for handling GRO_NORMAL skbs") > > Reported-by: Prijesh <prpatel@redhat.com> > > Does Prijesh have a last name? :) 'Prijesh <prpatel@redhat.com>' is what I got from his emails. But yes, he has. 'Prijesh Patel'. > > > Signed-off-by: Xin Long <lucien.xin@gmail.com> > > --- > > Documentation/admin-guide/sysctl/net.rst | 10 ++++++++++ > > 1 file changed, 10 insertions(+) > > > > diff --git a/Documentation/admin-guide/sysctl/net.rst b/Documentation/admin-guide/sysctl/net.rst > > index f86b5e1623c6..d8a8506f31ad 100644 > > --- a/Documentation/admin-guide/sysctl/net.rst > > +++ b/Documentation/admin-guide/sysctl/net.rst > > @@ -374,6 +374,16 @@ option is set to SOCK_TXREHASH_DEFAULT (i. e. not overridden by setsockopt). > > If set to 1 (default), hash rethink is performed on listening socket. > > If set to 0, hash rethink is not performed. > > > > +gro_normal_batch > > +---------------- > > + > > +Maximum number of GRO_NORMAL skbs to batch up for list-RX. When GRO decides > > +not to coalesce a packet, instead of passing it to the stack immediately, > > +place it on a list. > > That makes it sounds like only packets which were not coalesced > go on the list. IIUC everything goes on that list before traveling > up the stack, no? I think the difference is these ones held/merged go to gro_list first and get merged there, then go to the list. I can change it to: "place it on a list where the coalesced packets also eventually go" looks good? > > > +Pass this list to the stack at flush time or whenever > > This sentences is in second person, and the previous one was in third > person. how about "the list will be passed to stack at flush time..." ? > > > +the number of skbs in this list exceeds gro_normal_batch. > > s/skbs/segments/ right. Thanks. > > > +Default : 8 > > + > > 2. /proc/sys/net/unix - Parameters for Unix domain sockets > > ---------------------------------------------------------- > > >
On 10/05/2022 18:10, Xin Long wrote: > On Mon, May 9, 2022 at 9:07 PM Jakub Kicinski <kuba@kernel.org> wrote: >> That makes it sounds like only packets which were not coalesced >> go on the list. IIUC everything goes on that list before traveling >> up the stack, no? > I think the difference is these ones held/merged go to gro_list first > and get merged there, then go to the list. I can change it to: > > "place it on a list where the coalesced packets also eventually go" > > looks good? Maybe it'd be clearer to say something like "when a packet exits GRO, either as a coalesced superframe or as an original packet which GRO has decided not to coalesce, it is placed on a per-NAPI list. This list is then passed to the stack when..." etc. Ideally also mention the fact that a coalesced superframe counts as napi_gro_cb.count towards the gro_normal_batch limit, not just 1. -ed
On Tue, May 10, 2022 at 1:42 PM Edward Cree <edward.cree@xilinx.com> wrote: > > On 10/05/2022 18:10, Xin Long wrote: > > On Mon, May 9, 2022 at 9:07 PM Jakub Kicinski <kuba@kernel.org> wrote: > >> That makes it sounds like only packets which were not coalesced > >> go on the list. IIUC everything goes on that list before traveling > >> up the stack, no? > > I think the difference is these ones held/merged go to gro_list first > > and get merged there, then go to the list. I can change it to: > > > > "place it on a list where the coalesced packets also eventually go" > > > > looks good? > > Maybe it'd be clearer to say something like > "when a packet exits GRO, either as a coalesced superframe or as an > original packet which GRO has decided not to coalesce, it is placed on > a per-NAPI list. This list is then passed to the stack when..." etc. > Ideally also mention the fact that a coalesced superframe counts as > napi_gro_cb.count towards the gro_normal_batch limit, not just 1. Thanks, ed. I'd say "...when the segments in this list count towards the gro_normal_batch limit." only (too many details involved may confuse users) will post v2.
diff --git a/Documentation/admin-guide/sysctl/net.rst b/Documentation/admin-guide/sysctl/net.rst index f86b5e1623c6..d8a8506f31ad 100644 --- a/Documentation/admin-guide/sysctl/net.rst +++ b/Documentation/admin-guide/sysctl/net.rst @@ -374,6 +374,16 @@ option is set to SOCK_TXREHASH_DEFAULT (i. e. not overridden by setsockopt). If set to 1 (default), hash rethink is performed on listening socket. If set to 0, hash rethink is not performed. +gro_normal_batch +---------------- + +Maximum number of GRO_NORMAL skbs to batch up for list-RX. When GRO decides +not to coalesce a packet, instead of passing it to the stack immediately, +place it on a list. Pass this list to the stack at flush time or whenever +the number of skbs in this list exceeds gro_normal_batch. + +Default : 8 + 2. /proc/sys/net/unix - Parameters for Unix domain sockets ----------------------------------------------------------
Describe it in admin-guide/sysctl/net.rst like other Network core options. Users need to know gro_normal_batch for performance tuning. Fixes: 323ebb61e32b ("net: use listified RX for handling GRO_NORMAL skbs") Reported-by: Prijesh <prpatel@redhat.com> Signed-off-by: Xin Long <lucien.xin@gmail.com> --- Documentation/admin-guide/sysctl/net.rst | 10 ++++++++++ 1 file changed, 10 insertions(+)