Re: Softirq priority inversion from "softirq: reduce latencies"

From: Peter Hurley
Date: Sat Feb 27 2016 - 21:10:51 EST

On 02/27/2016 05:59 PM, Eric Dumazet wrote:
> On sam., 2016-02-27 at 15:33 -0800, Peter Hurley wrote:
>> On 02/27/2016 03:04 PM, David Miller wrote:
>>> From: Peter Hurley <peter@xxxxxxxxxxxxxxxxxx>
>>> Date: Sat, 27 Feb 2016 12:29:39 -0800
>>>> Not really. softirq raised from interrupt context will always execute
>>>> on this cpu and not in ksoftirqd, unless load forces softirq loop abort.
>>> That guarantee never was specified.
>> ??
>> Neither is running network socket servers at normal priority as if they're
>> higher priority than softirq.
>>> Or are you saying that by design, on a system under load, your UART
>>> will not function properly?
>>> Surely you don't mean that.
>> No, that's not what I mean.
>> What I mean is that bypassing the entire SOFTIRQ priority so that
>> sshd can process one network packet makes a mockery of the point of softirq.
>> This hack to workaround NET_RX looping over-and-over-and-over affects every
>> subsystem, not just one uart.
>> HI, TIMER, BLOCK; all of these are skipped: that's straight-up, a bug.
> No idea what you talk about.
> All pending softirq interrupts are processed. _Nothing_ is skipped.

An interrupt that schedules HI softirq while in NET_RX softirq should
still run the HI softirq. But with your patch that won't happen.

> Really, your system stability seems to depend on a completely
> undocumented behavior of linux kernels before linux-3.8
> If I understood, you expect that a tasklet activated from a softirq
> handler is run from the same __do_softirq() loop. This never has been
> the case.


The *interrupt handler* for DMA goes off while NET_RX softirq is running.
That's what schedules the *DMA tasklet*.

That tasklet should run before any process.

But it doesn't because your patch bails out early from softirq.

> My change simply triggers the bug in your driver earlier. As David
> pointed out, your bug should trigger the same on a loaded machine, even
> if you revert my patch.
> I honestly do not know why you arm a tasklet from NET_RX, why don't you
> simply process this directly, so that you do not rely on some scheduler
> decision ?