| Manual section: | 8 |
|---|
TCP has a number of buffer size tunables; auto-tuning is provided for them.
net.ipv4.tcp_wmem is a triple of min, default, max. By instrumenting tcp_sndbuf_expand() we can see where expansion is close to hitting the max, and we can adjust it up appropriately to allow for additional buffer space. This will not be done for cases where memory is low, and if we approach memory exhaustion and cannot increase overall tcp memory exhaustion limit (see below), the wmem max value will be decreased.
Similarly, for net.ipv4.tcp_rmem, we monitor and increase the limit when expansion is close to hitting the limit, with the same exceptions applying as above.
In both cases, we want to avoid the situation that increasing these limits leads to TCP memory exhaustion. The BPF programs that detect approach to those limits will not request increases if we are close to either TCP memory pressure or TCP memory exhaustion; in fact wmem/rmem will be reduced as part of an effort to decrease TCP memory usage if TCP memory exhaustion is approached and that value cannot be raised.
Similarly we want to avoid the other negative consequence of allocating too many buffers - latencies due to waiting to send/receive with longer queues. A blocking sender app that sends a lot of traffic will see less ENOBUFS errors and silently dropped packets, while a non-blocking app will see less EAGAIN messages. In those cases, facilitating sending will always be quicker and critically lead to a reduction in overall numbers of system calls. Similarly, a latency- sensitive app will likely prefer sends to succeed than have to retry. From the receive side, we have to consider the effect of a larger receive queue (and receive window). By default, we advertise half of the recieve buffer size as receive window; this allows for apps to use the rest as buffer space. This ratio of app space to window size can be adjusted via the sysctl tcp_adv_win_scale, which defaults to 1. A negative value means the receive window is scaled by the factor specified; so -2 means 1/4 of recieve buffer size is available for TCP window. A positive value of 2 means that 3/4 of receive buffer size is available for the TCP window.
So for slow apps, a negative value might make sense.
net.ipv4.tcp_mem represents the min, pressure, max values for overall TCP memory use in pages.
When in TCP memory pressure mode, we reclaim socket memory more aggressively until we fall below the tcp_mem min value. We reclaim the forward-allocated memory for example. On startup, TCP mem values are initialized as ~4.6%, 6.25% and 9.37% of nr_free_buffer_pages(). nr_free_buffer_pages() counts the number of pages beyond the high watermark in ZONE_DMA and ZONE_NORMAL.
As with watermark scaling, if we enter TCP memory pressure, bpftune will scale up min/pressure/max as needed, with limits of 6%/9% on min, pressure and 25% of available memory for the memory exhaustion max. We attempt to avoid memory exhaustion where possible, but if we hit the limit of memory exhaustion and cannot increase it further, wmem and rmem max values are decreased to reduce per-socket overhead.