Re: [tip:irq/core] genirq: Handle force threading of irqs with primary and thread handler

[Felipe Balbi]

Hi Thomas,

tip-bot for Thomas Gleixner <tipbot@xxxxxxxxx> writes:
> Commit-ID:  2a1d3ab8986d1b2f598ffc42351d94166fa0f022
> Gitweb:     http://git.kernel.org/tip/2a1d3ab8986d1b2f598ffc42351d94166fa0f022
> Author:     Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> AuthorDate: Mon, 21 Sep 2015 11:01:10 +0200
> Committer:  Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> CommitDate: Tue, 22 Sep 2015 12:39:57 +0200
>
> genirq: Handle force threading of irqs with primary and thread handler
>
> Force threading of interrupts does not really deal with interrupts
> which are requested with a primary and a threaded handler. The current
> policy is to leave them alone and let the primary handler run in
> interrupt context, but we set the ONESHOT flag for those interrupts as
> well.
>
> Kohji Okuno debugged a problem with the SDHCI driver where the
> interrupt thread waits for a hardware interrupt to trigger, which can't
> work well because the hardware interrupt is masked due to the ONESHOT
> flag being set. He proposed to set the ONESHOT flag only if the
> interrupt does not provide a thread handler.
>
> Though that does not work either because these interrupts can be
> shared. So the other interrupt would rightfully get the ONESHOT flag
> set and therefor the same situation would happen again.
>
> To deal with this proper, we need to force thread the primary handler
> of such interrupts as well. That means that the primary interrupt
> handler is treated as any other primary interrupt handler which is not
> marked IRQF_NO_THREAD. The threaded handler becomes a separate thread
> so the SDHCI flow logic can be handled gracefully.
>
> The same issue was reported against 4.1-rt.
>
> Reported-and-tested-by: Kohji Okuno <okuno.kohji@xxxxxxxxxxxxxxxx>
> Reported-By: Michal Smucr <msmucr@xxxxxxxxx>
> Reported-and-tested-by: Nathan Sullivan <nathan.sullivan@xxxxxx>
> Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> Link: http://lkml.kernel.org/r/alpine.DEB.2.11.1509211058080.5606@nanos
> Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>

this commit causes a performance regression for the USB driver on
several platforms (anybody using drivers/usb/dwc3, basically).

Here's the USB throughput with linux-next in 3 different scenarios:

1) Linux next without threadirqs cmdline

   test  0: sent     256.00 MB read      33.02 MB/s write      30.01 MB/s

2) Linux next with threadirqs on cmdline

   test  0: sent     256.00 MB read      30.70 MB/s write      27.89 MB/s

3) Linux next with threadirqs on cmdline + revert of $subject

   test  0: sent     256.00 MB read      32.93 MB/s write      29.85 MB/s


Considering this is trying to solve an issue found on the SDHCI driver,
shouldn't that be fixed instead ? Another option would be, of course, to
add IRQF_NO_THREAD to dwc3, but I'd like to avoid that if possible.

The way we try to use dwc3 is rather simple, actually. We use the
primary handle *only* to detect is $this device generated the IRQ and if
did we wake up the thread. We also don't make use of ONESHOT because we
mask $this device IRQs in the primary handler and only unmask after the
thread runs.

It's a bit surprising, to me at least, that simply running everything as
a thread would have such a measurable impact, but it does.

-- 
balbi

Attachment: signature.asc
Description: PGP signature