[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH] e1000e: Don't zero out buffer address in rx des

From: Kevin Wolf
Subject: Re: [Qemu-devel] [PATCH] e1000e: Don't zero out buffer address in rx descriptor
Date: Wed, 19 Oct 2016 12:07:40 +0200
User-agent: Mutt/1.5.21 (2010-09-15)

Am 19.10.2016 um 09:57 hat Dmitry Fleytman geschrieben:
>     On 19 Oct 2016, at 10:25 AM, Kevin Wolf <address@hidden> wrote:
>     Am 19.10.2016 um 08:48 hat Dmitry Fleytman geschrieben:
>            Another related thing that I noticed while debugging this and
>         turning on
>            tracing is that the interrupt throttling timers kept firing even if
>            there was no activity at all. Something might be wrong, there, too.
>            Next thing I wondered why throttling was enabled at all because the
>         spec
>            says the default is 0 (turned off). So one thing that I'm pretty
>         sure is
>            just a misunderstanding is the following defintion:
>            #define E1000E_MIN_XITR     (500) /* No more then 7813 interrupts
>         per
>                                                second according to spec
> */
>            As I understand it, the spec is just giving an example there and
>         lower
>            values are valid as well. At the very least, 0 should be accepted 
> as
>         a
>            special case because it means "disabled" and it's specified to be
>         the
>            default.
>         Right, this according to the spec this value should be 0 by default 
> and
>         throttling should be disabled.
>         Current device implementation does not allow specification of
>         throttling interval less than 500 and treats interval 0 as throttling
>         enabled with interval 500.
>         This is done by intention because according to the spec (
>         device cannot produce more than 7813 interrupts per second even when
>         throttling is disabled. Therefore, even in case of interrupt storm
>         (continuous interrupt re-injection by device), number of interrupts
>         produced by device is limited and CPU (driver) has enough time to do
>         its job and handle problematic interrupt state.
>     I think you're misinterpreting the spec here. This is the paragraph
>     we're talking about, right?
>        For example, if the interval is programmed to 500 (decimal), the
>        82574 guarantees the CPU is not interrupted by it for 128 µs from
>        the last interrupt. The maximum observable interrupt rate from the
>        82574 should never exceed 7813 interrupts/sec.
>     It says "for example", so this is just demonstrating how you can
>     calculate the effects of a specific throttling setting. It says that
>     _if_ you set ITR to 500, you get an interrupt at most every
>     500 * 256 ns = 128 µs. And 1 / 128 µs = 7821.5 Hz, so this is the
>     effective maximum frequency that _this specific_ ITR setting allows.
>     I also don't think it would make any sense for hardware to be unable to
>     trigger interrupts more often than that. Triggering an interrupt is not
>     a complex operation that involves a lot of calculation or anything.
> Hi Kevin,
> Yes, I assume that sentence
> “The maximum observable interrupt rate from the
> 82574 should never exceed 7813 interrupts/sec."
> is not a related to a specific case, but describes a generic limitation,
> however it might be I’m misreading the spec indeed.

For me everything hints at this being only an example: Not only do the
numbers match the example made in the previous sentence (which is
explicitly called an example) and look weird as a real limit, but it's
also in the same paragraph as the explicit example and the spec is
generally good at starting a new paragraph when talking about a new

I don't care enough to actually make you change anything, but I wanted
you to be aware that the interpretation of the spec as coded into our
emulation isn't clear at all (in fact, I would think it's clear that
it's _not_ meant this way) and that real hardware probably doesn't do
the same thing as we do.

What we're doing may still have merit, as a workaround for a guest
driver bug.

>         Opposed to this, virtual device is able to raise interrupts with rate
>         limited by CPU speed only therefore driver has no chance to fix
>         interrupt storm condition.
>         Windows e1000e drivers rely on upper limit for number of interrupts
>         per second in some cases and absence of this limit leads to infinite
>         interrupt storms.
>         To summarise, while usage of throttling mechanisms is a little bit
>         different from what specification says, effective emulated device
>         behavior is totally compliant to the real device.
>     So Windows doesn't configure ITR (i.e. it is 0) even though it can't
>     handle unlimited interrupts? That would be a driver bug then, and
>     perhaps an important enough one to keep a workaround in our code. But
>     then let's be explicit that this is a workaround for a Windows bug and
>     not mandated by the spec.
>     I'm not sure in what setup you produced this error, but possibly a
>     reason why this doesn't happen with real hardware isn't the NIC itself
>     but the backend: Communication with the host can obviously be faster
>     than talking to a physical network (so if you were doing the latter, the
>     rate in the VM wouldn't be limited by the CPU, but by the physical
>     network).
> This issue is reproduced on device disable and not related 
> to intensive device/backend communication. One RX packet with
> right timing is enough to trigged the problem.
> The same issue was fixed in e1000 device some time ago as well.

Commit 9596ef7c was good in flagging it as a guest driver bug. Only a
later series brought in the questionable spec interpretation.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]