Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 18 Mar 2021 10:39:35 -0600
From:   Alex Williamson <alex.williamson@redhat.com>
To:     Leon Romanovsky <leon@kernel.org>
Cc:     Amey Narkhede <ameynarkhede03@gmail.com>,
        raphael.norwitz@nutanix.com, linux-pci@vger.kernel.org,
        bhelgaas@google.com, linux-kernel@vger.kernel.org,
        alay.shah@nutanix.com, suresh.gumpula@nutanix.com,
        shyam.rajendran@nutanix.com, felipe@nutanix.com
Subject: Re: [PATCH 4/4] PCI/sysfs: Allow userspace to query and set device
 reset mechanism
Message-ID: <20210318103935.2ec32302@omen.home.shazbot.org>
In-Reply-To: <YFMYzkg101isRXIM@unreal>
References: <20210315102950.230de1d6@x1.home.shazbot.org>
        <20210315183226.GA14801@raphael-debian-dev>
        <YFGDgqdTLBhQL8mN@unreal>
        <20210317102447.73no7mhox75xetlf@archlinux>
        <YFHh3bopQo/CRepV@unreal>
        <20210317112309.nborigwfd26px2mj@archlinux>
        <YFHsW/1MF6ZSm8I2@unreal>
        <20210317131718.3uz7zxnvoofpunng@archlinux>
        <YFILEOQBOLgOy3cy@unreal>
        <20210317113140.3de56d6c@omen.home.shazbot.org>
        <YFMYzkg101isRXIM@unreal>
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, 18 Mar 2021 11:09:34 +0200
Leon Romanovsky <leon@kernel.org> wrote:

> On Wed, Mar 17, 2021 at 11:31:40AM -0600, Alex Williamson wrote:
> > On Wed, 17 Mar 2021 15:58:40 +0200
> > Leon Romanovsky <leon@kernel.org> wrote:
> > =20
> > > On Wed, Mar 17, 2021 at 06:47:18PM +0530, Amey Narkhede wrote: =20
> > > > On 21/03/17 01:47PM, Leon Romanovsky wrote: =20
> > > > > On Wed, Mar 17, 2021 at 04:53:09PM +0530, Amey Narkhede wrote: =20
> > > > > > On 21/03/17 01:02PM, Leon Romanovsky wrote: =20
> > > > > > > On Wed, Mar 17, 2021 at 03:54:47PM +0530, Amey Narkhede wrote=
: =20
> > > > > > > > On 21/03/17 06:20AM, Leon Romanovsky wrote: =20
> > > > > > > > > On Mon, Mar 15, 2021 at 06:32:32PM +0000, Raphael Norwitz=
 wrote: =20
> > > > > > > > > > On Mon, Mar 15, 2021 at 10:29:50AM -0600, Alex Williams=
on wrote: =20
> > > > > > > > > > > On Mon, 15 Mar 2021 21:03:41 +0530
> > > > > > > > > > > Amey Narkhede <ameynarkhede03@gmail.com> wrote:
> > > > > > > > > > > =20
> > > > > > > > > > > > On 21/03/15 05:07PM, Leon Romanovsky wrote: =20
> > > > > > > > > > > > > On Mon, Mar 15, 2021 at 08:34:09AM -0600, Alex Wi=
lliamson wrote: =20
> > > > > > > > > > > > > > On Mon, 15 Mar 2021 14:52:26 +0100
> > > > > > > > > > > > > > Pali Roh=C3=A1r <pali@kernel.org> wrote:
> > > > > > > > > > > > > > =20
> > > > > > > > > > > > > > > On Monday 15 March 2021 19:13:23 Amey Narkhed=
e wrote: =20
> > > > > > > > > > > > > > > > slot reset (pci_dev_reset_slot_function) an=
d secondary bus
> > > > > > > > > > > > > > > > reset(pci_parent_bus_reset) which I think a=
re hot reset and
> > > > > > > > > > > > > > > > warm reset respectively. =20
> > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > No. PCI secondary bus reset =3D PCIe Hot Rese=
t. Slot reset is just another
> > > > > > > > > > > > > > > type of reset, which is currently implemented=
 only for PCIe hot plug
> > > > > > > > > > > > > > > bridges and for PowerPC PowerNV platform and =
it just call PCI secondary
> > > > > > > > > > > > > > > bus reset with some other hook. PCIe Warm Res=
et does not have API in
> > > > > > > > > > > > > > > kernel and therefore drivers do not export th=
is type of reset via any
> > > > > > > > > > > > > > > kernel function (yet). =20
> > > > > > > > > > > > > >
> > > > > > > > > > > > > > Warm reset is beyond the scope of this series, =
but could be implemented
> > > > > > > > > > > > > > in a compatible way to fit within the pci_reset=
_fn_methods[] array
> > > > > > > > > > > > > > defined here.  Note that with this series the r=
esets available through
> > > > > > > > > > > > > > pci_reset_function() and the per device reset a=
ttribute is sysfs remain
> > > > > > > > > > > > > > exactly the same as they are currently.  The bu=
s and slot reset
> > > > > > > > > > > > > > methods used here are limited to devices where =
only a single function is
> > > > > > > > > > > > > > affected by the reset, therefore it is not like=
 the patch you proposed
> > > > > > > > > > > > > > which performed a reset irrespective of the dow=
nstream devices.  This
> > > > > > > > > > > > > > series only enables selection of the existing m=
ethods.  Thanks, =20
> > > > > > > > > > > > >
> > > > > > > > > > > > > Alex,
> > > > > > > > > > > > >
> > > > > > > > > > > > > I asked the patch author here [1], but didn't get=
 any response, maybe
> > > > > > > > > > > > > you can answer me. What is the use case scenario =
for this functionality?
> > > > > > > > > > > > >
> > > > > > > > > > > > > Thanks
> > > > > > > > > > > > >
> > > > > > > > > > > > > [1] https://lore.kernel.org/lkml/YE389lAqjJSeTolM=
@unreal/
> > > > > > > > > > > > > =20
> > > > > > > > > > > > Sorry for not responding immediately. There were so=
me buggy wifi cards
> > > > > > > > > > > > which needed FLR explicitly not sure if that behavi=
or is fixed in
> > > > > > > > > > > > drivers. Also there is use a case at Nutanix but th=
e engineer who
> > > > > > > > > > > > is involved is on PTO that is why I did not respond=
 immediately as
> > > > > > > > > > > > I don't know the details yet. =20
> > > > > > > > > > >
> > > > > > > > > > > And more generally, devices continue to have reset is=
sues and we
> > > > > > > > > > > impose a fixed priority in our ordering.  We can and =
probably should
> > > > > > > > > > > continue to quirk devices when we find broken resets =
so that we have
> > > > > > > > > > > the best default behavior, but it's currently not eas=
y for an end user
> > > > > > > > > > > to experiment, ie. this reset works, that one doesn't=
.  We might also
> > > > > > > > > > > have platform issues where a given reset works better=
 on a certain
> > > > > > > > > > > platform.  Exposing a way to test these things might =
lead to better
> > > > > > > > > > > quirks.  In the case I think Pali was looking for, th=
ey wanted a
> > > > > > > > > > > mechanism to force a bus reset, if this was in refere=
nce to a single
> > > > > > > > > > > function device, this could be accomplished by settin=
g a priority for
> > > > > > > > > > > that mechanism, which would translate to not only the=
 sysfs reset
> > > > > > > > > > > attribute, but also the reset mechanism used by vfio-=
pci.  Thanks,
> > > > > > > > > > >
> > > > > > > > > > > Alex
> > > > > > > > > > > =20
> > > > > > > > > >
> > > > > > > > > > To confirm from our end - we have seen many such instan=
ces where default
> > > > > > > > > > reset methods have not worked well on our platform. Deb=
ugging these
> > > > > > > > > > issues is painful in practice, and this interface would=
 make it far
> > > > > > > > > > easier.
> > > > > > > > > >
> > > > > > > > > > Having an interface like this would also help us better=
 communicate the
> > > > > > > > > > issues we find with upstream. Allowing others to more e=
asily test our
> > > > > > > > > > (or other entities') findings should give better visibi=
lity into
> > > > > > > > > > which issues apply to the device in general and which a=
re platform
> > > > > > > > > > specific. In disambiguating the former from the latter,=
 we should be
> > > > > > > > > > able to better quirk devices for everyone, and in the l=
atter cases, this
> > > > > > > > > > interface allows for a safer and more elegant solution =
than any of the
> > > > > > > > > > current alternatives. =20
> > > > > > > > >
> > > > > > > > > So to summarize, we are talking about test and debug inte=
rface to
> > > > > > > > > overcome HW bugs, am I right?
> > > > > > > > >
> > > > > > > > > My personal experience shows that once the easy workaroun=
d exists
> > > > > > > > > (and write to generally available sysfs is very simple), =
the vendors
> > > > > > > > > and users desire for proper fix decreases drastically. IM=
HO, we will
> > > > > > > > > see increase of copy/paste in SO and blog posts, but redu=
ce in quirks.
> > > > > > > > >
> > > > > > > > > My 2-cents.
> > > > > > > > > =20
> > > > > > > > I agree with your point but at least it gives the userspace=
 ability
> > > > > > > > to use broken device until bug is fixed in upstream. =20
> > > > > > >
> > > > > > > As I said, I don't expect many fixes once "userspace" will be=
 able to
> > > > > > > use cheap workaround. There is no incentive to fix it. =20
> >
> > We can increase the annoyance factor of using a modified set of reset
> > methods, but ultimately we can only control what goes into our kernel,
> > other kernels might take v1 of this series and incorporate it
> > regardless of what happens here.
> > =20
> > > > > > > > This is also applicable for obscure devices without upstream
> > > > > > > > drivers for example custom FPGA based devices. =20
> > > > > > >
> > > > > > > This is not relevant to upstream kernel. Those vendors ship e=
verything
> > > > > > > custom, they don't need upstream, we don't need them :)
> > > > > > > =20
> > > > > > By custom I meant hobbyists who could tinker with their custom =
FPGA. =20
> > > > >
> > > > > I invite such hobbyists to send patches and include their FPGA in
> > > > > upstream kernel. =20
> >
> > This is potentially another good use case, how receptive are we going
> > to be to an FPGA design that botches a reset.  Do they have a valid
> > device ID for us to base a quirk on, are they just squatting on one, or
> > using the default from a library.  Maybe the next bitstream will
> > resolve it, maybe without any external indication.  IOW, what would the
> > quality level be for that quirk versus using this as a workaround,
> > where the user probably wouldn't mind a kernel nag? =20
>=20
> It is worth to solve it when the need arises.
>=20
> > =20
> > > > > > > > Another main application which I forgot to mention is virtu=
alization
> > > > > > > > where vmm wants to reset the device when the guest is reset,
> > > > > > > > to emulate machine reboot as closely as possible. =20
> > > > > > >
> > > > > > > It can work in very narrow case, because reset will cause to =
device
> > > > > > > reprobe and most likely the driver will be different from the=
 one that
> > > > > > > started reset. I can imagine that net devices will lose their=
 state and
> > > > > > > config after such reset too.
> > > > > > > =20
> > > > > > Not sure if I got that 100% right. The pci_reset_function() fun=
ction
> > > > > > saves and restores device state over the reset. =20
> > > > >
> > > > > I'm talking about netdev state, but whatever given the existence =
of
> > > > > sysfs reset knob.
> > > > > =20
> > > > > > =20
> > > > > > > IMHO, it will be saner for everyone if virtualization don't t=
ry such resets. =20
> >
> > That would cause a massive regression in device assignment support.  As
> > with other sysfs attributes, triggering them alongside a running driver
> > is probably not going to end well.  However, pci_reset_function() is
> > extremely useful for stopping devices and returning them to a default
> > state, when either rebooting a VM or returning the device to the host.
> > The device is not removed and re-probed when this occurs, vfio-pci is
> > able to hold onto the device across these actions.  Sure, don't reset a
> > netdev device when it's in use, that's not what these are used for.
> > =20
> > > > > > The exists reset sysfs attribute was added for exactly this case
> > > > > > though. =20
> > > > >
> > > > > I didn't know the rationale behind that file till you said and I
> > > > > googled libvirt discussion, so ok. Do you propose that libvirt
> > > > > will manage database of devices and their working reset types?
> > > > > =20
> > > > I don't have much idea about internals of libvirt but why would
> > > > it need to manage database of working reset types? It could just
> > > > read new reset_methods attribute to get the list of supported reset
> > > > methods. =20
> > >
> > > Because the idea of this patch is to read all supported reset types a=
nd
> > > allow to the user to chose the working one. The user will do it with
> > > help from StackOverflow, but libvirt will need to have some sort of
> > > database, otherwise it won't be different from simple "echo 1 > reset"
> > > which will iterate over all supported resets anyway. =20
> >
> > AFAIK, libvirt no longer attempts to do resets itself, or is at least
> > moving in that direction.  vfio-pci will reset as device when they're
> > opened by a user (when available) or triggered via the API. =20
>=20
> <...>
>=20
> > > The difference here is that this is a workaround to solve bugs that
> > > should be fixed in the kernel. =20
> >
> > If we want to discourage using this as a primary means to resolve reset
> > issues on a device then we can create log warnings any time it's used.
> > Downstreams that really want this functionality are going to take this
> > patch from the list whether we accept it or not.  As above, it seems
> > there are valid use cases.  Even with mainstream vfio in QEMU, I go
> > through some hoops trying to determine if I can do a secondary bus
> > reset rather than a PM reset because it's not specified anywhere what a
> > "soft reset" means for any given device.  This sort of interface could
> > make it easier to apply a system policy that a pci_reset_function()
> > should always perform a secondary bus reset if the only other option is
> > a PM reset.  Maybe that policy mostly makes sense for a VM use case, so
> > we'd want one policy by default and another when the device is used for
> > this functionality.  How could we accomplish that with a quirk?  Thanks=
, =20
>=20
> I'm lost here, does vfio-pci use sysfs interface or internal to the kerne=
l API?
>=20
> If it is latter then we don't really need sysfs, if not, we still need
> some sort of DB to create second policy, because "supported !=3D working".
> What am I missing?

vfio-pci uses the internal kernel API, ie. the variants of
pci_reset_function(), which is the same interface used by the existing
sysfs reset mechanism.  This proposed configuration of the reset method
would affect any driver using that same core infrastructure and from my
perspective that's really the goal.  In the case where a supported
reset mechanism fails for a device, continuing to quirk those out for
the best default behavior makes sense, I'd be disappointed for a vendor
to not pursue improving the default behavior where it clearly makes
sense.  However, there's also a policy decision, the kernel imposes a
preferential ordering of reset mechanism.  Is that ordering the best
case for all users?  I've presented above a case where a userspace may
prefer a policy of preferring a bus reset to a PM reset.  So I think
the question is not only are there supported mechanisms that don't
work, where this interface allows userspace to more readily identify
and work around those sorts of issues, but it also enables user
preference and easier evaluation whether all of the supported reset
mechanisms work rather than just the first one we encounter in the
ordering we've decided to impose today.  Thanks,

Alex