On Wed, Feb 16, 2022 at 2:25 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote: > > On Tue, Feb 15, 2022 at 9:03 PM <kerneljasonxing@xxxxxxxxx> wrote: > > > > From: Jason Xing <xingwanli@xxxxxxxxxxxx> > > > > Normally, user doesn't care the logic behind the kernel if they're > > trying to set receive buffer via setsockopt. However, once the new > > value of the receive buffer is set even though it's not smaller than > > the initial value which is sysctl_tcp_rmem[1] implemented in > > tcp_rcv_space_adjust(),, the server's wscale will shrink and then > > lead to the bad bandwidth as intended. > > Quite confusing changelog, honestly. > > Users of SO_RCVBUF specifically told the kernel : I want to use _this_ > buffer size, I do not want the kernel to decide for me. > > Also, I think your changelog does not really explain that _if_ you set > SO_RCVBUF to a small value before > connect() or in general the 3WHS, the chosen wscale will be small, and > this won't allow future 10x increase > of the effective RWIN. > Yes, you hit the point really. > > > > > For now, introducing a new socket option to let the receive buffer > > grow automatically no matter what the new value is can solve > > the bad bandwidth issue meanwhile it's not breaking the application > > with SO_RCVBUF option set. > > > > Here are some numbers: > > $ sysctl -a | grep rmem > > net.core.rmem_default = 212992 > > net.core.rmem_max = 40880000 > > net.ipv4.tcp_rmem = 4096 425984 40880000 > > > > Case 1 > > on the server side > > # iperf -s -p 5201 > > on the client side > > # iperf -c [client ip] -p 5201 > > It turns out that the bandwidth is 9.34 Gbits/sec while the wscale of > > server side is 10. It's good. > > > > Case 2 > > on the server side > > #iperf -s -p 5201 -w 425984 > > on the client side > > # iperf -c [client ip] -p 5201 > > It turns out that the bandwidth is reduced to 2.73 Gbits/sec while the > > wcale is 2, even though the receive buffer is not changed at all at the > > very beginning. > > > > After this patch is applied, the bandwidth of case 2 is recovered to > > 9.34 Gbits/sec as expected at the cost of consuming more memory per > > socket. > > How does your patch allow wscale to increase after flow is established ? > > I would remove from the changelog these experimental numbers that look > quite wrong, > maybe copy/pasted from your prior version. > My fault. I should have removed this part. > Instead I would describe why an application might want to clear the > 'receive buffer size is locked' socket attribute. > > > > > Signed-off-by: Jason Xing <xingwanli@xxxxxxxxxxxx> > > -- > > v2: suggested by Eric > > - introduce new socket option instead of breaking the logic in SO_RCVBUF > > - Adjust the title and description of this patch > > link: https://lore.kernel.org/lkml/CANn89iL8vOUOH9bZaiA-cKcms+PotuKCxv7LpVx3RF0dDDSnmg@xxxxxxxxxxxxxx/ > > --- > > > > I think adding another parallel SO_RCVBUF option is not good. It is > adding confusion (and net/core/filter.c has been unchanged) I'll change the filter.c altogether in the next submission. > > Also we want CRIU to work correctly. > > So if you have a SO_XXXX setsockopt() call, you also need to provide > getsockopt() implementation. > > I would suggest an option to clear or set SOCK_RCVBUF_LOCK, and > getsockopt() would return if the bit is currently set or not. > > Something clearly describing the intent, like SO_RCVBUF_LOCK maybe. Just now, I found out that the latest kernel has merged a similar patch (commit 04190bf89) about three months ago. Is it still necessary to add another separate option to clear the SOCK_RCVBUF_LOCK explicitly? Thanks, Jason
