Difference in buffer behaviour between 8.89 and 8.9?

[rapier]

Honestly, I don't know. It shouldn't grow any larger than buf->max_size 
which seems to have a hard limit of 134,217,728 bytes. That said, in 
memory constrained systems the TCP receive buffer probably isn't going 
to be large enough for this really to make a difference.

Ideally I'd like to set the growth of the buffer to be a function of the 
tcp receive buffer. I am getting that information in hpnssh but getting 
the extern from channels.c to sshbuf.c is causing a weird link error in 
sshkey.c. I haven't nailed that down yet.

Note: resetting the CHAN_SES_DEFAULT_WINDOW to the size in the patch is 
just to demonstrate the problem. If you aren't making that buffer bigger 
than it's current size (of 2MB) this should never be an issue.

It is an issue for my work in hpnssh because we dynamically adjust SSH's 
receive buffer to match auto-tuning TCP receive buffers in order to 
maximize throughput.

The main thing is that there was a change in how the buffers are handled 
between 8.8 and 8.9 that could, in some cases, end up impacting 
performance.

On 5/19/22 2:37 PM, Thorsten Glaser wrote:
> On Thu, 19 May 2022, rapier wrote:
> 
>> then aggressively grow that buffer (4MB at a time)
> 
> Does this impact low-memory systems like my SPARCstation
> and that old 80486 laptop where growing this much may fail?
> 
> bye,
> //mirabilos