Re: [PATCH 1/3] PCI/ASPM: Use the path max in L1 ASPM latency check

From: Bjorn Helgaas
Date: Mon Dec 14 2020 - 09:03:26 EST

On Mon, Dec 14, 2020 at 10:14:18AM +0100, Ian Kumlien wrote:
> On Mon, Dec 14, 2020 at 6:44 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> >
> > [+cc Jesse, Tony, David, Jakub, Heiner, lists in case there's an ASPM
> > issue with I211 or Realtek NICs. Beginning of thread:
> > https://lore.kernel.org/r/20201024205548.1837770-1-ian.kumlien@xxxxxxxxx
> >
> > Short story: Ian has:
> >
> > Root Port --- Switch --- I211 NIC
> > \-- multifunction Realtek NIC, etc
> >
> > and the I211 performance is poor with ASPM L1 enabled on both links
> > in the path to it. The patch here disables ASPM on the upstream link
> > and fixes the performance, but AFAICT the devices in that path give us
> > no reason to disable L1. If I understand the spec correctly, the
> > Realtek device should not be relevant to the I211 path.]
> >
> > On Sun, Dec 13, 2020 at 10:39:53PM +0100, Ian Kumlien wrote:
> > > On Sun, Dec 13, 2020 at 12:47 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > > > On Sat, Oct 24, 2020 at 10:55:46PM +0200, Ian Kumlien wrote:
> > > > > Make pcie_aspm_check_latency comply with the PCIe spec, specifically:
> > > > > "5.4.1.2.2. Exit from the L1 State"
> > > > >
> > > > > Which makes it clear that each switch is required to
> > > > > initiate a transition within 1μs from receiving it,
> > > > > accumulating this latency and then we have to wait for the
> > > > > slowest link along the path before entering L0 state from
> > > > > L1.
> > > > > ...
> > > >
> > > > > On my specific system:
> > > > > 03:00.0 Ethernet controller: Intel Corporation I211 Gigabit Network Connection (rev 03)
> > > > > 04:00.0 Unassigned class [ff00]: Realtek Semiconductor Co., Ltd. Device 816e (rev 1a)
> > > > >
> > > > > Exit latency Acceptable latency
> > > > > Tree: L1 L0s L1 L0s
> > > > > ---------- ------- ----- ------- ------
> > > > > 00:01.2 <32 us -
> > > > > | 01:00.0 <32 us -
> > > > > |- 02:03.0 <32 us -
> > > > > | \03:00.0 <16 us <2us <64 us <512ns
> > > > > |
> > > > > \- 02:04.0 <32 us -
> > > > > \04:00.0 <64 us unlimited <64 us <512ns
> > > > >
> > > > > 04:00.0's latency is the same as the maximum it allows so as
> > > > > we walk the path the first switchs startup latency will pass
> > > > > the acceptable latency limit for the link, and as a
> > > > > side-effect it fixes my issues with 03:00.0.
> > > > >
> > > > > Without this patch, 03:00.0 misbehaves and only gives me ~40
> > > > > mbit/s over links with 6 or more hops. With this patch I'm
> > > > > back to a maximum of ~933 mbit/s.
> > > >
> > > > There are two paths here that share a Link:
> > > >
> > > > 00:01.2 --- 01:00.0 -- 02:03.0 --- 03:00.0 I211 NIC
> > > > 00:01.2 --- 01:00.0 -- 02:04.0 --- 04:00.x multifunction Realtek
> > > >
> > > > 1) The path to the I211 NIC includes four Ports and two Links (the
> > > > connection between 01:00.0 and 02:03.0 is internal Switch routing,
> > > > not a Link).
> > >
> > > > The Ports advertise L1 exit latencies of <32us, <32us, <32us,
> > > > <16us. If both Links are in L1 and 03:00.0 initiates L1 exit at T,
> > > > 01:00.0 initiates L1 exit at T + 1. A TLP from 03:00.0 may see up
> > > > to 1 + 32 = 33us of L1 exit latency.
> > > >
> > > > The NIC can tolerate up to 64us of L1 exit latency, so it is safe
> > > > to enable L1 for both Links.
> > > >
> > > > 2) The path to the Realtek device is similar except that the Realtek
> > > > L1 exit latency is <64us. If both Links are in L1 and 04:00.x
> > > > initiates L1 exit at T, 01:00.0 again initiates L1 exit at T + 1,
> > > > but a TLP from 04:00.x may see up to 1 + 64 = 65us of L1 exit
> > > > latency.
> > > >
> > > > The Realtek device can only tolerate 64us of latency, so it is not
> > > > safe to enable L1 for both Links. It should be safe to enable L1
> > > > on the shared link because the exit latency for that link would be
> > > > <32us.
> > >
> > > 04:00.0:
> > > DevCap: MaxPayload 128 bytes, PhantFunc 0, Latency L0s <512ns, L1 <64us
> > > LnkCap: Port #0, Speed 5GT/s, Width x1, ASPM L0s L1, Exit Latency L0s
> > > unlimited, L1 <64us
> > >
> > > So maximum latency for the entire link has to be <64 us
> > > For the device to leave L1 ASPM takes <64us
> > >
> > > So the device itself is the slowest entry along the link, which
> > > means that nothing else along that path can have ASPM enabled
> >
> > Yes. That's what I said above: "it is not safe to enable L1 for both
> > Links." Unless I'm missing something, we agree on that.
> >
> > I also said that it should be safe to enable L1 on the shared Link
> > (from 00:01.2 to 01:00.0) because if the downstream Link is always in
> > L0, the exit latency of the shared Link should be <32us, and 04:00.x
> > can tolerate 64us.
>
> Exit latency of shared link would be max of link, ie 64 + L1-hops, not 32

I don't think this is true. The path from 00:01.2 to 04:00.x includes
two Links, and they are independent. The exit latency for each Link
depends only on the Port at each end:

Link 1 (depends on 00:01.2 and 01:00.0): max(32, 32) = 32us
Link 2 (depends on 02:04.0 and 04:00.x): max(32, 64) = 64us

If L1 is enabled for Link 1 and disabled for Link 2, Link 2 will
remain in L0 so it has no L1 exit latency, and the exit latency of
the entire path should be 32us.

> > > > > The original code path did:
> > > > > 04:00:0-02:04.0 max latency 64 -> ok
> > > > > 02:04.0-01:00.0 max latency 32 +1 -> ok
> > > > > 01:00.0-00:01.2 max latency 32 +2 -> ok
> > > > >
> > > > > And thus didn't see any L1 ASPM latency issues.
> > > > >
> > > > > The new code does:
> > > > > 04:00:0-02:04.0 max latency 64 -> ok
> > > > > 02:04.0-01:00.0 max latency 64 +1 -> latency exceeded
> > > > > 01:00.0-00:01.2 max latency 64 +2 -> latency exceeded
> > > >
> > > > [Nit: I don't think we should add 1 for the 02:04.0 -- 01:00.0 piece
> > > > because that's internal Switch routing, not a Link. But even without
> > > > that extra microsecond, this path does exceed the acceptable latency
> > > > since 1 + 64 = 65us, and 04:00.0 can only tolerate 64us.]
> > >
> > > It does report L1 ASPM on both ends, so the links will be counted as
> > > such in the code.
> >
> > This is a bit of a tangent and we shouldn't get too wrapped up in it.
> > This is a confusing aspect of PCIe. We're talking about this path:
> >
> > 00:01.2 --- [01:00.0 -- 02:04.0] --- 04:00.x multifunction Realtek
> >
> > This path only contains two Links. The first one is
> > 00:01.2 --- 01:00.0, and the second one is 02:04.0 --- 04:00.x.
> >
> > 01:00.0 is a Switch Upstream Port and 02:04.0 is a Switch Downstream
> > Port. The connection between them is not a Link; it is some internal
> > wiring of the Switch that is completely opaque to software.
> >
> > The ASPM information and knobs in 01:00.0 apply to the Link on its
> > upstream side, and the ASPM info and knobs in 02:04.0 apply to the
> > Link on its downstream side.
> >
> > The example in sec 5.4.1.2.2 contains three Links. The L1 exit latency
> > for the Link is the max of the exit latencies at each end:
> >
> > Link 1: max(32, 8) = 32us
> > Link 2: max(8, 32) = 32us
> > Link 3: max(32, 8) = 32us
> >
> > The total delay for a TLP starting at the downstream end of Link 3
> > is 32 + 2 = 32us.
> >
> > In the path to your 04:00.x Realtek device:
> >
> > Link 1 (from 00:01.2 to 01:00.0): max(32, 32) = 32us
> > Link 2 (from 02:04.0 to 04:00.x): max(32, 64) = 64us
> >
> > If L1 were enabled on both Links, the exit latency would be 64 + 1 =
> > 65us.
>
> So one line to be removed from the changelog, i assume... And yes, the
> code handles that - first disable is 01:00.0 <-> 00:01.2
>
> > > I also assume that it can power down individual ports... and enter
> > > rest state if no links are up.
> >
> > I don't think this is quite true -- a Link can't enter L1 unless the
> > Ports on both ends have L1 enabled, so I don't think it makes sense to
> > talk about an individual Port being in L1.
> >
> > > > > It correctly identifies the issue.
> > > > >
> > > > > For reference, pcie information:
> > > > > https://bugzilla.kernel.org/show_bug.cgi?id=209725
> > > >
> > > > The "lspci without my patches" [1] shows L1 enabled for the shared
> > > > Link from 00:01.2 --- 01:00.0 and for the Link to 03:00.0 (I211), but
> > > > not for the Link to 04:00.x (Realtek).
> > > >
> > > > Per my analysis above, that looks like it *should* be a safe
> > > > configuration. 03:00.0 can tolerate 64us, actual is <33us. 04:00.0
> > > > can tolerate 64us, actual should be <32us since only the shared Link
> > > > is in L1.
> > >
> > > See above.
> >
> > As I said above, if we enabled L1 only on the shared Link from 00:01.2
> > to 01:00.0, the exit latency should be acceptable. In that case, a
> > TLP from 04:00.x would see only 32us of latency:
> >
> > Link 1 (from 00:01.2 to 01:00.0): max(32, 32) = 32us
> >
> > and 04:00.x can tolerate 64us.
>
> But, again, you're completely ignoring the full link, ie 04:00.x would
> also have to power on.

I think you're using "the full link" to refer to the entire path from
00:01.2 to 04:00.x. In PCIe, a "Link" directly connects two Ports.
It doesn't refer to the entire path.

No, if L1 is disabled on 02:04.0 and 04:00.x (as Linux apparently does
by default), the Link between them never enters L1, so there is no
power-on for this Link.

> > > > However, the commit log at [2] shows L1 *enabled* for both the shared
> > > > Link from 00:01.2 --- 01:00.0 and the 02:04.0 --- 04:00.x Link, and
> > > > that would definitely be a problem.
> > > >
> > > > Can you explain the differences between [1] and [2]?
> > >
> > > I don't understand which sections you're referring to.
> >
> > [1] is the "lspci without my patches" attachment of bugzilla #209725,
> > which is supposed to show the problem this patch solves. We're
> > talking about the path to 04:00.x, and [1] show this:
> >
> > 01:00.2 L1+ # <-- my typo here, should be 00:01.2
> > 01:00.0 L1+
> > 02:04.0 L1-
> > 04:00.0 L1-
> >
> > AFAICT, that should be a legal configuration as far as 04:00.0 is
> > concerned, so it's not a reason for this patch.
>
> Actually, no, maximum path latency 64us
>
> 04:00.0 wakeup latency == 64us
>
> Again, as stated, it can't be behind any sleeping L1 links

It would be pointless for a device to advertise L1 support if it could
never be used. 04:00.0 advertises that it can tolerate L1 latency of
64us and that it can exit L1 in 64us or less. So it *can* be behind a
Link in L1 as long as nothing else in the path adds more latency.

> > [2] is a previous posting of this same patch, and its commit log
> > includes information about the same path to 04:00.x, but the "LnkCtl
> > Before" column shows:
> >
> > 01:00.2 L1+ # <-- my typo here, should be 00:01.2
> > 01:00.0 L1+
> > 02:04.0 L1+
> > 04:00.0 L1+
> >
> > I don't know why [1] shows L1 disabled on the downstream Link, while
> > [2] shows L1 *enabled* on the same Link.
>
> From the data they look switched.
>
> > > > > Kai-Heng Feng has a machine that will not boot with ASPM without
> > > > > this patch, information is documented here:
> > > > > https://bugzilla.kernel.org/show_bug.cgi?id=209671
> > > >
> > > > I started working through this info, too, but there's not enough
> > > > information to tell what difference this patch makes. The attachments
> > > > compare:
> > > >
> > > > 1) CONFIG_PCIEASPM_DEFAULT=y without the patch [3] and
> > > > 2) CONFIG_PCIEASPM_POWERSAVE=y *with* the patch [4]
> > > >
> > > > Obviously CONFIG_PCIEASPM_POWERSAVE=y will configure things
> > > > differently than CONFIG_PCIEASPM_DEFAULT=y, so we can't tell what
> > > > changes are due to the config change and what are due to the patch.
> > > >
> > > > The lspci *with* the patch ([4]) shows L0s and L1 enabled at almost
> > > > every possible place. Here are the Links, how they're configured, and
> > > > my analysis of the exit latencies vs acceptable latencies:
> > > >
> > > > 00:01.1 --- 01:00.0 L1+ ( L1 <64us vs unl)
> > > > 00:01.2 --- 02:00.0 L1+ ( L1 <64us vs 64us)
> > > > 00:01.3 --- 03:00.0 L1+ ( L1 <64us vs 64us)
> > > > 00:01.4 --- 04:00.0 L1+ ( L1 <64us vs unl)
> > > > 00:08.1 --- 05:00.x L0s+ L1+ (L0s <64ns vs 4us, L1 <1us vs unl)
> > > > 00:08.2 --- 06:00.0 L0s+ L1+ (L0s <64ns vs 4us, L1 <1us vs unl)
> > > >
> > > > So I can't tell what change prevents the freeze. I would expect the
> > > > patch would cause us to *disable* L0s or L1 somewhere.
> > > >
> > > > The only place [4] shows ASPM disabled is for 05:00.1. The spec says
> > > > we should program the same value in all functions of a multi-function
> > > > device. This is a non-ARI device, so "only capabilities enabled in
> > > > all functions are enabled for the component as a whole." That would
> > > > mean that L0s and L1 are effectively disabled for 05:00.x even though
> > > > 05:00.0 claims they're enabled. But the latencies say ASPM L0s and L1
> > > > should be safe to be enabled. This looks like another bug that's
> > > > probably unrelated.
> > >
> > > I don't think it's unrelated, i suspect it's how PCIe works with
> > > multiple links... a device can cause some kind of head of queue
> > > stalling - i don't know how but it really looks like it.
> >
> > The text in quotes above is straight out of the spec (PCIe r5.0, sec
> > 7.5.3.7). Either the device works that way or it's not compliant.
> >
> > The OS configures ASPM based on the requirements and capabilities
> > advertised by the device. If a device has any head of queue stalling
> > or similar issues, those must be comprehended in the numbers
> > advertised by the device. It's not up to the OS to speculate about
> > issues like that.
> >
> > > > The patch might be correct; I haven't actually analyzed the code. But
> > > > the commit log doesn't make sense to me yet.
> > >
> > > I personally don't think that all this PCI information is required,
> > > the linux kernel is currently doing it wrong according to the spec.
> >
> > We're trying to establish exactly *what* Linux is doing wrong. So far
> > we don't have a good explanation of that.
>
> Yes we do, linux counts hops + max for "link" while what should be done is
> counting hops + max for path

I think you're saying we need to include L1 exit latency even for
Links where L1 is disabled. I don't think we should include those.

> > Based on [1], in the path to 03:00.0, both Links have L1 enabled, with
> > an exit latency of <33us, and 03:00.0 can tolerate 64us. That should
> > work fine.
> >
> > Also based on [1], in the path to 04:00.x, the upstream Link has L1
> > enabled and the downstream Link has L1 disabled, for an exit latency
> > of <32us, and 04:00.0 can tolerate 64us. That should also work fine.
>
> Again, ignoring the exit latendy for 04:00.0
>
> > (Alternately, disabling L1 on the upstream Link and enabling it on the
> > downstream Link should have an exit latency of <64us and 04:00.0 can
> > tolerate 64us, so that should work fine, too.)
>
> Then nothing else can have L1 aspm enabled

Yes, as I said, we should be able to enable L1 on either of the Links
in the path to 04:00.x, but not both.

The original problem here is not with the Realtek device at 04:00.x
but with the I211 NIC at 03:00.0. So we also need to figure out what
the connection is. Does the same I211 performance problem occur if
you remove the Realtek device from the system?

03:00.0 can tolerate 64us of latency, so even if L1 is enabled on both
Links leading to it, the path exit latency would be <33us, which
should be fine.

> > > Also, since it's clearly doing the wrong thing, I'm worried that
> > > dists will take a kernel enable aspm and there will be alot of
> > > bugreports of non-booting systems or other weird issues... And the
> > > culprit was known all along.
> >
> > There's clearly a problem on your system, but I don't know yet whether
> > Linux is doing something wrong, a device in your system is designed
> > incorrectly, or a device is designed correctly but the instance in
> > your system is defective.
>
> According to the spec it is, there is a explanation of how to
> calculate the exit latency
> and when you implement that, which i did (before knowing the actual
> spec) then it works...
>
> > > It's been five months...
> >
> > I apologize for the delay. ASPM is a subtle area of PCIe, the Linux
> > code is complicated, and we have a long history of issues with it. I
> > want to fix the problem, but I want to make sure we do it in a way
> > that matches the spec so the fix applies to all systems. I don't want
> > a magic fix that fixes your system in a way I don't quite understand.
>
> > Obviously *you* understand this, so hopefully it's just a matter of
> > pounding it through my thick skull :)
>
> I only understand what I've been forced to understand - and I do
> leverage the existing code without
> knowing what it does underneath, I only look at the links maximum
> latency and make sure that I keep
> the maximum latency along the path and not just link for link
>
> once you realise that the max allowed latency is buffer dependent -
> then this becomes obviously correct,
> and then the pcie spec showed it as being correct as well... so...
>
>
> > > > [1] https://bugzilla.kernel.org/attachment.cgi?id=293047
> > > > [2] https://lore.kernel.org/linux-pci/20201007132808.647589-1-ian.kumlien@xxxxxxxxx/
> > > > [3] https://bugzilla.kernel.org/attachment.cgi?id=292955
> > > > [4] https://bugzilla.kernel.org/attachment.cgi?id=292957
> > > >
> > > > > Signed-off-by: Ian Kumlien <ian.kumlien@xxxxxxxxx>
> > > > > Tested-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
> > > > > ---
> > > > > drivers/pci/pcie/aspm.c | 22 ++++++++++++++--------
> > > > > 1 file changed, 14 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> > > > > index 253c30cc1967..c03ead0f1013 100644
> > > > > --- a/drivers/pci/pcie/aspm.c
> > > > > +++ b/drivers/pci/pcie/aspm.c
> > > > > @@ -434,7 +434,7 @@ static void pcie_get_aspm_reg(struct pci_dev *pdev,
> > > > >
> > > > > static void pcie_aspm_check_latency(struct pci_dev *endpoint)
> > > > > {
> > > > > - u32 latency, l1_switch_latency = 0;
> > > > > + u32 latency, l1_max_latency = 0, l1_switch_latency = 0;
> > > > > struct aspm_latency *acceptable;
> > > > > struct pcie_link_state *link;
> > > > >
> > > > > @@ -456,10 +456,14 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
> > > > > if ((link->aspm_capable & ASPM_STATE_L0S_DW) &&
> > > > > (link->latency_dw.l0s > acceptable->l0s))
> > > > > link->aspm_capable &= ~ASPM_STATE_L0S_DW;
> > > > > +
> > > > > /*
> > > > > * Check L1 latency.
> > > > > - * Every switch on the path to root complex need 1
> > > > > - * more microsecond for L1. Spec doesn't mention L0s.
> > > > > + *
> > > > > + * PCIe r5.0, sec 5.4.1.2.2 states:
> > > > > + * A Switch is required to initiate an L1 exit transition on its
> > > > > + * Upstream Port Link after no more than 1 μs from the beginning of an
> > > > > + * L1 exit transition on any of its Downstream Port Links.
> > > > > *
> > > > > * The exit latencies for L1 substates are not advertised
> > > > > * by a device. Since the spec also doesn't mention a way
> > > > > @@ -469,11 +473,13 @@ static void pcie_aspm_check_latency(struct pci_dev *endpoint)
> > > > > * L1 exit latencies advertised by a device include L1
> > > > > * substate latencies (and hence do not do any check).
> > > > > */
> > > > > - latency = max_t(u32, link->latency_up.l1, link->latency_dw.l1);
> > > > > - if ((link->aspm_capable & ASPM_STATE_L1) &&
> > > > > - (latency + l1_switch_latency > acceptable->l1))
> > > > > - link->aspm_capable &= ~ASPM_STATE_L1;
> > > > > - l1_switch_latency += 1000;
> > > > > + if (link->aspm_capable & ASPM_STATE_L1) {
> > > > > + latency = max_t(u32, link->latency_up.l1, link->latency_dw.l1);
> > > > > + l1_max_latency = max_t(u32, latency, l1_max_latency);
> > > > > + if (l1_max_latency + l1_switch_latency > acceptable->l1)
> > > > > + link->aspm_capable &= ~ASPM_STATE_L1;
> > > > > + l1_switch_latency += 1000;
> > > > > + }
> > > > >
> > > > > link = link->parent;
> > > > > }
> > > > > --
> > > > > 2.29.1
> > > > >