net: Introduce preferred busy-polling
authorBjörn Töpel <bjorn.topel@intel.com>
Mon, 30 Nov 2020 18:51:56 +0000 (19:51 +0100)
committerDaniel Borkmann <daniel@iogearbox.net>
Mon, 30 Nov 2020 23:09:25 +0000 (00:09 +0100)
commit7fd3253a7de6a317a0683f83739479fb880bffc8
tree8490a9756c6a4a7f5078fd856f23b568703e3680
parent854055c0cf30d732b3514ce7956976f60496b1a1
net: Introduce preferred busy-polling

The existing busy-polling mode, enabled by the SO_BUSY_POLL socket
option or system-wide using the /proc/sys/net/core/busy_read knob, is
an opportunistic. That means that if the NAPI context is not
scheduled, it will poll it. If, after busy-polling, the budget is
exceeded the busy-polling logic will schedule the NAPI onto the
regular softirq handling.

One implication of the behavior above is that a busy/heavy loaded NAPI
context will never enter/allow for busy-polling. Some applications
prefer that most NAPI processing would be done by busy-polling.

This series adds a new socket option, SO_PREFER_BUSY_POLL, that works
in concert with the napi_defer_hard_irqs and gro_flush_timeout
knobs. The napi_defer_hard_irqs and gro_flush_timeout knobs were
introduced in commit 6f8b12d661d0 ("net: napi: add hard irqs deferral
feature"), and allows for a user to defer interrupts to be enabled and
instead schedule the NAPI context from a watchdog timer. When a user
enables the SO_PREFER_BUSY_POLL, again with the other knobs enabled,
and the NAPI context is being processed by a softirq, the softirq NAPI
processing will exit early to allow the busy-polling to be performed.

If the application stops performing busy-polling via a system call,
the watchdog timer defined by gro_flush_timeout will timeout, and
regular softirq handling will resume.

In summary; Heavy traffic applications that prefer busy-polling over
softirq processing should use this option.

Example usage:

  $ echo 2 | sudo tee /sys/class/net/ens785f1/napi_defer_hard_irqs
  $ echo 200000 | sudo tee /sys/class/net/ens785f1/gro_flush_timeout

Note that the timeout should be larger than the userspace processing
window, otherwise the watchdog will timeout and fall back to regular
softirq processing.

Enable the SO_BUSY_POLL/SO_PREFER_BUSY_POLL options on your socket.

Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Link: https://lore.kernel.org/bpf/20201130185205.196029-2-bjorn.topel@gmail.com
arch/alpha/include/uapi/asm/socket.h
arch/mips/include/uapi/asm/socket.h
arch/parisc/include/uapi/asm/socket.h
arch/sparc/include/uapi/asm/socket.h
fs/eventpoll.c
include/linux/netdevice.h
include/net/busy_poll.h
include/net/sock.h
include/uapi/asm-generic/socket.h
net/core/dev.c
net/core/sock.c