TCP receive windows broken in 4.14.48

TCP receive windows broken in 4.14.48

[Benjamin Gilbert]

Hi,

4.14.48 can cause abnormally small TCP receive windows when the sender is
faster than the receiver; see https://github.com/coreos/bugs/issues/2457 for
details.  Reverting "tcp: avoid integer overflows in tcp_rcv_space_adjust()"
fixes it.  Backporting 02db55718d53 ("tcp: do not overshoot window_clamp in
tcp_rcv_space_adjust()"), which is its parent commit upstream, also fixes
it.

--Benjamin Gilbert

Re: TCP receive windows broken in 4.14.48

[Guenter Roeck]

On Thu, Jun 14, 2018 at 05:09:21PM -0400, Benjamin Gilbert wrote:
> Hi,
> 
> 4.14.48 can cause abnormally small TCP receive windows when the sender is
> faster than the receiver; see https://github.com/coreos/bugs/issues/2457 for
> details.  Reverting "tcp: avoid integer overflows in tcp_rcv_space_adjust()"
> fixes it.  Backporting 02db55718d53 ("tcp: do not overshoot window_clamp in
> tcp_rcv_space_adjust()"), which is its parent commit upstream, also fixes
> it.
> 
02db55718d53 applies cleanly, so I would suggest we go along that route
instead of a revert.

Greg - given that the problematic commit is in all stable releases, it would
probably make sense to apply 02db55718d53 to pre-4.14 kernels as well. It
does apply clean to all kernels all the way back to 3.18.y.

Thanks,
Guenter

Re: TCP receive windows broken in 4.14.48

[Greg Kroah-Hartman]

On Thu, Jun 14, 2018 at 02:20:29PM -0700, Guenter Roeck wrote:
> On Thu, Jun 14, 2018 at 05:09:21PM -0400, Benjamin Gilbert wrote:
> > Hi,
> > 
> > 4.14.48 can cause abnormally small TCP receive windows when the sender is
> > faster than the receiver; see https://github.com/coreos/bugs/issues/2457 for
> > details.  Reverting "tcp: avoid integer overflows in tcp_rcv_space_adjust()"
> > fixes it.  Backporting 02db55718d53 ("tcp: do not overshoot window_clamp in
> > tcp_rcv_space_adjust()"), which is its parent commit upstream, also fixes
> > it.
> > 
> 02db55718d53 applies cleanly, so I would suggest we go along that route
> instead of a revert.
> 
> Greg - given that the problematic commit is in all stable releases, it would
> probably make sense to apply 02db55718d53 to pre-4.14 kernels as well. It
> does apply clean to all kernels all the way back to 3.18.y.

Now done, thanks.

greg k-h

Re: TCP receive windows broken in 4.14.48

[Greg Kroah-Hartman]

On Thu, Jun 14, 2018 at 02:20:29PM -0700, Guenter Roeck wrote:
> On Thu, Jun 14, 2018 at 05:09:21PM -0400, Benjamin Gilbert wrote:
> > Hi,
> > 
> > 4.14.48 can cause abnormally small TCP receive windows when the sender is
> > faster than the receiver; see https://github.com/coreos/bugs/issues/2457 for
> > details.  Reverting "tcp: avoid integer overflows in tcp_rcv_space_adjust()"
> > fixes it.  Backporting 02db55718d53 ("tcp: do not overshoot window_clamp in
> > tcp_rcv_space_adjust()"), which is its parent commit upstream, also fixes
> > it.
> > 
> 02db55718d53 applies cleanly, so I would suggest we go along that route
> instead of a revert.
> 
> Greg - given that the problematic commit is in all stable releases, it would
> probably make sense to apply 02db55718d53 to pre-4.14 kernels as well. It
> does apply clean to all kernels all the way back to 3.18.y.

Ah, but it breaks the build everywhere, so I've dropped it :(

Can someone please backport it?  I'll be glad to queue it up then.

thanks,

greg k-h

Re: TCP receive windows broken in 4.14.48

[Guenter Roeck]

On 06/17/2018 10:20 PM, Greg Kroah-Hartman wrote:
> On Thu, Jun 14, 2018 at 02:20:29PM -0700, Guenter Roeck wrote:
>> On Thu, Jun 14, 2018 at 05:09:21PM -0400, Benjamin Gilbert wrote:
>>> Hi,
>>>
>>> 4.14.48 can cause abnormally small TCP receive windows when the sender is
>>> faster than the receiver; see https://github.com/coreos/bugs/issues/2457 for
>>> details.  Reverting "tcp: avoid integer overflows in tcp_rcv_space_adjust()"
>>> fixes it.  Backporting 02db55718d53 ("tcp: do not overshoot window_clamp in
>>> tcp_rcv_space_adjust()"), which is its parent commit upstream, also fixes
>>> it.
>>>
>> 02db55718d53 applies cleanly, so I would suggest we go along that route
>> instead of a revert.
>>
>> Greg - given that the problematic commit is in all stable releases, it would
>> probably make sense to apply 02db55718d53 to pre-4.14 kernels as well. It
>> does apply clean to all kernels all the way back to 3.18.y.
> 
> Ah, but it breaks the build everywhere, so I've dropped it :(
> 

Sigh. Maybe I'll learn to compile test at some point.

> Can someone please backport it?  I'll be glad to queue it up then.
> 

Just sent a backport, though with a bit confused subject. I did compile
test it this time with 4.14.y and 4.9.y.

Guenter