Dear Linux folks,
On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
almost five seconds to initialize PCI. According to the timestamps, 1.5
s are from assigning the PCI devices to the 142 IOMMU groups.
```
$ lspci | wc -l
281
$ dmesg
[…]
[ 2.918411] PCI: Using host bridge windows from ACPI; if necessary,
use "pci=nocrs" and report a bug
[ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
[ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
[ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM
ClockPM Segments MSI HPX-Type3]
[ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
[ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME PCIeCapability]
[ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported,
using BIOS configuration
[ 3.010511] PCI host bridge to bus 0000:00
[…]
[ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
[ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02 (active)
[ 6.239906] pnp: PnP ACPI: found 6 devices
[…]
[ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device
[8086:2034]
[ 6.996063] PCI: CLS 0 bytes, default 64
[ 7.000008] Trying to unpack rootfs image as initramfs...
[ 7.065281] Freeing initrd memory: 5136K
[…]
[ 7.079098] DMAR: dmar7: Using Queued invalidation
[ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
[…]
[ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
[ 8.571191] DMAR: Intel(R) Virtualization Technology for Directed I/O
[ 8.577618] PCI-DMA: Using software bounce buffering for IO (SWIOTLB)
[…]
```
Is there anything that could be done to reduce the time?
Kind regards,
Paul
Dear Linux folks,
Am 05.11.21 um 12:56 schrieb Paul Menzel:
> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
> almost five seconds to initialize PCI. According to the timestamps, 1.5
> s are from assigning the PCI devices to the 142 IOMMU groups.
>
> ```
> $ lspci | wc -l
> 281
> $ dmesg
> […]
> [ 2.918411] PCI: Using host bridge windows from ACPI; if necessary, use "pci=nocrs" and report a bug
> [ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
> [ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
> [ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM ClockPM Segments MSI HPX-Type3]
> [ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
> [ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME PCIeCapability]
> [ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported,
> using BIOS configuration
> [ 3.010511] PCI host bridge to bus 0000:00
> […]
> [ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
> [ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02 (active)
> [ 6.239906] pnp: PnP ACPI: found 6 devices
> […]
> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device [8086:2034]
> [ 6.996063] PCI: CLS 0 bytes, default 64
> [ 7.000008] Trying to unpack rootfs image as initramfs...
> [ 7.065281] Freeing initrd memory: 5136K
> […]
> [ 7.079098] DMAR: dmar7: Using Queued invalidation
> [ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
> […]
> [ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
> [ 8.571191] DMAR: Intel(R) Virtualization Technology for Directed I/O
> [ 8.577618] PCI-DMA: Using software bounce buffering for IO (SWIOTLB)
> […]
> ```
>
> Is there anything that could be done to reduce the time?
I created an issue at the Kernel.org Bugzilla, and attached the output
of `dmesg` there [1].
Kind regards,
Paul
[1]: https://bugzilla.kernel.org/show_bug.cgi?id=214953
On Fri, Nov 05, 2021 at 12:56:09PM +0100, Paul Menzel wrote:
> Dear Linux folks,
>
>
> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
> almost five seconds to initialize PCI. According to the timestamps, 1.5 s
> are from assigning the PCI devices to the 142 IOMMU groups.
>
> ```
> $ lspci | wc -l
> 281
> $ dmesg
> […]
> [ 2.918411] PCI: Using host bridge windows from ACPI; if necessary, use
> "pci=nocrs" and report a bug
> [ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
> [ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
> [ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM
> ClockPM Segments MSI HPX-Type3]
> [ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
> [ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME PCIeCapability]
> [ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported, using
> BIOS configuration
> [ 3.010511] PCI host bridge to bus 0000:00
> […]
> [ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
> [ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02 (active)
> [ 6.239906] pnp: PnP ACPI: found 6 devices
For ~280 PCI devices, (6.24-2.92)/280 = 0.012 s/dev. On my laptop I
have about (.66-.37)/36 = 0.008 s/dev (on v5.4), so about the same
ballpark.
Faster would always be better, of course. I assume this is not really
a regression?
> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device
> [8086:2034]
> [ 6.996063] PCI: CLS 0 bytes, default 64
> [ 7.000008] Trying to unpack rootfs image as initramfs...
> [ 7.065281] Freeing initrd memory: 5136K
> […]
> [ 7.079098] DMAR: dmar7: Using Queued invalidation
> [ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
> […]
> [ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
I don't have this iommu stuff turned on and don't know what's
happening here.
> Is there anything that could be done to reduce the time?
>
>
> Kind regards,
>
> Paul
Dear Bjorn,
Thank you for your quick reply.
Am 05.11.21 um 19:53 schrieb Bjorn Helgaas:
> On Fri, Nov 05, 2021 at 12:56:09PM +0100, Paul Menzel wrote:
>> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
>> almost five seconds to initialize PCI. According to the timestamps, 1.5 s
>> are from assigning the PCI devices to the 142 IOMMU groups.
>>
>> ```
>> $ lspci | wc -l
>> 281
>> $ dmesg
>> […]
>> [ 2.918411] PCI: Using host bridge windows from ACPI; if necessary, use "pci=nocrs" and report a bug
>> [ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
>> [ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
>> [ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM ClockPM Segments MSI HPX-Type3]
>> [ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
>> [ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME PCIeCapability]
>> [ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported, using BIOS configuration
>> [ 3.010511] PCI host bridge to bus 0000:00
>> […]
>> [ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
>> [ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02 (active)
>> [ 6.239906] pnp: PnP ACPI: found 6 devices
>
> For ~280 PCI devices, (6.24-2.92)/280 = 0.012 s/dev. On my laptop I
> have about (.66-.37)/36 = 0.008 s/dev (on v5.4), so about the same
> ballpark.
Though if it was on average 0.008 s/dev here, around a second could be
saved.
The integrated Matrox G200eW3 graphics controller (102b:0536) and the
two Broadcom NetXtreme BCM5720 2-port Gigabit Ethernet PCIe cards
(14e4:165f) take 150 ms to be initialized.
[ 3.454409] pci 0000:03:00.0: [102b:0536] type 00 class 0x030000
[ 3.460411] pci 0000:03:00.0: reg 0x10: [mem
0x91000000-0x91ffffff pref]
[ 3.467403] pci 0000:03:00.0: reg 0x14: [mem 0x92808000-0x9280bfff]
[ 3.473402] pci 0000:03:00.0: reg 0x18: [mem 0x92000000-0x927fffff]
[ 3.479437] pci 0000:03:00.0: BAR 0: assigned to efifb
The timestamp in each line differs by around 6 ms. Could printing the
messages to the console (VGA) hold this up (line 373 to line 911 makes
(6.24 s-2.92 s)/(538 lines) = (3.32 s)/(538 lines) = 6 ms)?
[ 3.484480] pci 0000:02:00.0: PCI bridge to [bus 03]
[ 3.489401] pci 0000:02:00.0: bridge window [mem
0x92000000-0x928fffff]
[ 3.496398] pci 0000:02:00.0: bridge window [mem
0x91000000-0x91ffffff 64bit pref]
[ 3.504446] pci 0000:04:00.0: [14e4:165f] type 00 class 0x020000
[ 3.510415] pci 0000:04:00.0: reg 0x10: [mem
0x92e30000-0x92e3ffff 64bit pref]
[ 3.517408] pci 0000:04:00.0: reg 0x18: [mem
0x92e40000-0x92e4ffff 64bit pref]
[ 3.524407] pci 0000:04:00.0: reg 0x20: [mem
0x92e50000-0x92e5ffff 64bit pref]
[ 3.532402] pci 0000:04:00.0: reg 0x30: [mem
0xfffc0000-0xffffffff pref]
[ 3.538483] pci 0000:04:00.0: PME# supported from D0 D3hot D3cold
[ 3.544437] pci 0000:04:00.0: 4.000 Gb/s available PCIe
bandwidth, limited by 5.0 GT/s PCIe x1 link at 0000:00:1c.5 (capable of
8.000 Gb/s with 5.0 GT/s PCIe x2 link)
[ 3.559493] pci 0000:04:00.1: [14e4:165f] type 00 class 0x020000
Here is a 15 ms delay.
[ 3.565415] pci 0000:04:00.1: reg 0x10: [mem
0x92e00000-0x92e0ffff 64bit pref]
[ 3.573407] pci 0000:04:00.1: reg 0x18: [mem
0x92e10000-0x92e1ffff 64bit pref]
[ 3.580407] pci 0000:04:00.1: reg 0x20: [mem
0x92e20000-0x92e2ffff 64bit pref]
[ 3.587402] pci 0000:04:00.1: reg 0x30: [mem
0xfffc0000-0xffffffff pref]
[ 3.594483] pci 0000:04:00.1: PME# supported from D0 D3hot D3cold
[ 3.600502] pci 0000:00:1c.5: PCI bridge to [bus 04]
Can the 6 ms – also from your system – be explained by the PCI
specification? Seeing how fast PCI nowadays is, 6 ms sounds like a long
time. ;-)
> Faster would always be better, of course. I assume this is not really
> a regression?
Correct, as far as I know of, this is no regression.
>> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device [8086:2034]
>> [ 6.996063] PCI: CLS 0 bytes, default 64
>> [ 7.000008] Trying to unpack rootfs image as initramfs...
>> [ 7.065281] Freeing initrd memory: 5136K
The PCI resource assignment(?) also seems to take 670 ms:
[ 6.319656] pci 0000:04:00.0: can't claim BAR 6 [mem
0xfffc0000-0xffffffff pref]: no compatible bridge window
[…]
[ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device
[8086:2034]
>> […]
>> [ 7.079098] DMAR: dmar7: Using Queued invalidation
>> [ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
>> […]
>> [ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
>
> I don't have this iommu stuff turned on and don't know what's
> happening here.
There is a lock in `iommu_group_add_device()` in `drivers/iommu/iommu.c`:
mutex_lock(&group->mutex);
list_add_tail(&device->list, &group->devices);
if (group->domain && !iommu_is_attach_deferred(group->domain,
dev))
ret = __iommu_attach_device(group->domain, dev);
mutex_unlock(&group->mutex);
No idea, if it’s related. Unfortunately, it’s a production system, so I
can’t do any debugging. (Maybe `initcall_debug` could give some
insight.) Maybe the IOMMU developers can explain it without it. Could
the IOMMU group assignment be done in parallel?
>> Is there anything that could be done to reduce the time?
Kind regards,
Paul
Hi Paul,
> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
> almost five seconds to initialize PCI. According to the timestamps, 1.5 s
> are from assigning the PCI devices to the 142 IOMMU groups.
[...]
> Is there anything that could be done to reduce the time?
I am curious - why is this a problem? Are you power-cycling your servers
so often to the point where the cumulative time spent in enumerating PCI
devices and adding them later to IOMMU groups is a problem?
I am simply wondering why you decided to signal out the PCI enumeration as
slow in particular, especially given that a large server hardware tends to
have (most of the time, as per my experience) rather long initialisation
time either from being powered off or after being power cycled. I can take
a while before the actual operating system itself will start.
We talked about this briefly with Bjorn, and there might be an option to
perhaps add some caching, as we suspect that the culprit here is doing PCI
configuration space read for each device, which can be slow on some
platforms.
However, we would need to profile this to get some quantitative data to see
whether doing anything would even be worthwhile. It would definitely help
us understand better where the bottlenecks really are and of what magnitude.
I personally don't have access to such a large hardware like the one you
have access to, thus I was wondering whether you would have some time, and
be willing, to profile this for us on the hardware you have.
Let me know what do you think?
Krzysztof
On 2021-11-06 10:42, Paul Menzel wrote:
> Dear Bjorn,
>
>
> Thank you for your quick reply.
>
>
> Am 05.11.21 um 19:53 schrieb Bjorn Helgaas:
>> On Fri, Nov 05, 2021 at 12:56:09PM +0100, Paul Menzel wrote:
>
>>> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
>>> almost five seconds to initialize PCI. According to the timestamps,
>>> 1.5 s
>>> are from assigning the PCI devices to the 142 IOMMU groups.
>>>
>>> ```
>>> $ lspci | wc -l
>>> 281
>>> $ dmesg
>>> […]
>>> [ 2.918411] PCI: Using host bridge windows from ACPI; if
>>> necessary, use "pci=nocrs" and report a bug
>>> [ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
>>> [ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
>>> [ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig
>>> ASPM ClockPM Segments MSI HPX-Type3]
>>> [ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
>>> [ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME
>>> PCIeCapability]
>>> [ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported,
>>> using BIOS configuration
>>> [ 3.010511] PCI host bridge to bus 0000:00
>>> […]
>>> [ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
>>> [ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02
>>> (active)
>>> [ 6.239906] pnp: PnP ACPI: found 6 devices
>>
>> For ~280 PCI devices, (6.24-2.92)/280 = 0.012 s/dev. On my laptop I
>> have about (.66-.37)/36 = 0.008 s/dev (on v5.4), so about the same
>> ballpark.
>
> Though if it was on average 0.008 s/dev here, around a second could be
> saved.
>
> The integrated Matrox G200eW3 graphics controller (102b:0536) and the
> two Broadcom NetXtreme BCM5720 2-port Gigabit Ethernet PCIe cards
> (14e4:165f) take 150 ms to be initialized.
>
> [ 3.454409] pci 0000:03:00.0: [102b:0536] type 00 class 0x030000
> [ 3.460411] pci 0000:03:00.0: reg 0x10: [mem
> 0x91000000-0x91ffffff pref]
> [ 3.467403] pci 0000:03:00.0: reg 0x14: [mem 0x92808000-0x9280bfff]
> [ 3.473402] pci 0000:03:00.0: reg 0x18: [mem 0x92000000-0x927fffff]
> [ 3.479437] pci 0000:03:00.0: BAR 0: assigned to efifb
>
> The timestamp in each line differs by around 6 ms. Could printing the
> messages to the console (VGA) hold this up (line 373 to line 911 makes
> (6.24 s-2.92 s)/(538 lines) = (3.32 s)/(538 lines) = 6 ms)?
>
> [ 3.484480] pci 0000:02:00.0: PCI bridge to [bus 03]
> [ 3.489401] pci 0000:02:00.0: bridge window [mem
> 0x92000000-0x928fffff]
> [ 3.496398] pci 0000:02:00.0: bridge window [mem
> 0x91000000-0x91ffffff 64bit pref]
> [ 3.504446] pci 0000:04:00.0: [14e4:165f] type 00 class 0x020000
> [ 3.510415] pci 0000:04:00.0: reg 0x10: [mem
> 0x92e30000-0x92e3ffff 64bit pref]
> [ 3.517408] pci 0000:04:00.0: reg 0x18: [mem
> 0x92e40000-0x92e4ffff 64bit pref]
> [ 3.524407] pci 0000:04:00.0: reg 0x20: [mem
> 0x92e50000-0x92e5ffff 64bit pref]
> [ 3.532402] pci 0000:04:00.0: reg 0x30: [mem
> 0xfffc0000-0xffffffff pref]
> [ 3.538483] pci 0000:04:00.0: PME# supported from D0 D3hot D3cold
> [ 3.544437] pci 0000:04:00.0: 4.000 Gb/s available PCIe
> bandwidth, limited by 5.0 GT/s PCIe x1 link at 0000:00:1c.5 (capable of
> 8.000 Gb/s with 5.0 GT/s PCIe x2 link)
> [ 3.559493] pci 0000:04:00.1: [14e4:165f] type 00 class 0x020000
>
> Here is a 15 ms delay.
>
> [ 3.565415] pci 0000:04:00.1: reg 0x10: [mem
> 0x92e00000-0x92e0ffff 64bit pref]
> [ 3.573407] pci 0000:04:00.1: reg 0x18: [mem
> 0x92e10000-0x92e1ffff 64bit pref]
> [ 3.580407] pci 0000:04:00.1: reg 0x20: [mem
> 0x92e20000-0x92e2ffff 64bit pref]
> [ 3.587402] pci 0000:04:00.1: reg 0x30: [mem
> 0xfffc0000-0xffffffff pref]
> [ 3.594483] pci 0000:04:00.1: PME# supported from D0 D3hot D3cold
> [ 3.600502] pci 0000:00:1c.5: PCI bridge to [bus 04]
>
> Can the 6 ms – also from your system – be explained by the PCI
> specification? Seeing how fast PCI nowadays is, 6 ms sounds like a long
> time. ;-)
>
>> Faster would always be better, of course. I assume this is not really
>> a regression?
>
> Correct, as far as I know of, this is no regression.
>
>>> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device
>>> [8086:2034]
>>> [ 6.996063] PCI: CLS 0 bytes, default 64
>>> [ 7.000008] Trying to unpack rootfs image as initramfs...
>>> [ 7.065281] Freeing initrd memory: 5136K
>
> The PCI resource assignment(?) also seems to take 670 ms:
>
> [ 6.319656] pci 0000:04:00.0: can't claim BAR 6 [mem
> 0xfffc0000-0xffffffff pref]: no compatible bridge window
> […]
> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device
> [8086:2034]
>
>>> […]
>>> [ 7.079098] DMAR: dmar7: Using Queued invalidation
>>> [ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
>>> […]
>>> [ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
>>
>> I don't have this iommu stuff turned on and don't know what's
>> happening here.
>
> There is a lock in `iommu_group_add_device()` in `drivers/iommu/iommu.c`:
>
> mutex_lock(&group->mutex);
> list_add_tail(&device->list, &group->devices);
> if (group->domain && !iommu_is_attach_deferred(group->domain,
> dev))
> ret = __iommu_attach_device(group->domain, dev);
> mutex_unlock(&group->mutex);
>
> No idea, if it’s related. Unfortunately, it’s a production system, so I
> can’t do any debugging. (Maybe `initcall_debug` could give some
> insight.) Maybe the IOMMU developers can explain it without it. Could
> the IOMMU group assignment be done in parallel?
FWIW I'd expect that locking to be pretty much immaterial - many devices
are getting their own uncontended groups, and callers of this tend to be
serialised at a higher level anyway. iommu_probe_device() usually runs
off the back of the device_add() notifier (where it could be that it's
the only thing making noise in between something *else* being slow), but
there is the special case where it gets replayed for all existing
devices when the IOMMU driver registers itself - at a guess it seems
like it may well be the latter case you're seeing, but either way
there's not much to say without figuring out where the time is actually
being spent (I don't suppose that machine has dynamic ftrace enabled?).
That said, setting up a new group isn't a completely insignificant
amount of work, and 142 groups seems a lot - I'd have assumed that a
system of that scale would be the kind of big server kit that takes
several minutes to boot to the point of even starting the kernel anyway.
Robin.
>>> Is there anything that could be done to reduce the time?
>
>
> Kind regards,
>
> Paul
> _______________________________________________
> iommu mailing list
> [email protected]
> https://lists.linuxfoundation.org/mailman/listinfo/iommu
Dear Krzysztof,
Thank you for your reply.
Am 08.11.21 um 18:18 schrieb Krzysztof Wilczyński:
>> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
>> almost five seconds to initialize PCI. According to the timestamps, 1.5 s
>> are from assigning the PCI devices to the 142 IOMMU groups.
> [...]
>> Is there anything that could be done to reduce the time?
>
> I am curious - why is this a problem? Are you power-cycling your servers
> so often to the point where the cumulative time spent in enumerating PCI
> devices and adding them later to IOMMU groups is a problem?
>
> I am simply wondering why you decided to signal out the PCI enumeration as
> slow in particular, especially given that a large server hardware tends to
> have (most of the time, as per my experience) rather long initialisation
> time either from being powered off or after being power cycled. I can take
> a while before the actual operating system itself will start.
It’s not a problem per se, and more a pet peeve of mine. Systems get
faster and faster, and boottime slower and slower. On desktop systems,
it’s much more important with firmware like coreboot taking less than
one second to initialize the hardware and passing control to the
payload/operating system. If we are lucky, we are going to have servers
with FLOSS firmware.
But, already now, using kexec to reboot a system, avoids the problems
you pointed out on servers, and being able to reboot a system as quickly
as possible, lowers the bar for people to reboot systems more often to,
for example, so updates take effect.
> We talked about this briefly with Bjorn, and there might be an option to
> perhaps add some caching, as we suspect that the culprit here is doing PCI
> configuration space read for each device, which can be slow on some
> platforms.
>
> However, we would need to profile this to get some quantitative data to see
> whether doing anything would even be worthwhile. It would definitely help
> us understand better where the bottlenecks really are and of what magnitude.
>
> I personally don't have access to such a large hardware like the one you
> have access to, thus I was wondering whether you would have some time, and
> be willing, to profile this for us on the hardware you have.
>
> Let me know what do you think?
Sounds good. I’d be willing to help. Note, that I won’t have time before
Wednesday next week though.
Kind regards,
Paul
Dear Robin,
Thank you for your reply.
Am 09.11.21 um 16:31 schrieb Robin Murphy:
> On 2021-11-06 10:42, Paul Menzel wrote:
>> Am 05.11.21 um 19:53 schrieb Bjorn Helgaas:
>>> On Fri, Nov 05, 2021 at 12:56:09PM +0100, Paul Menzel wrote:
>>
>>>> On a PowerEdge T440/021KCD, BIOS 2.11.2 04/22/2021, Linux 5.10.70 takes
>>>> almost five seconds to initialize PCI. According to the timestamps,
>>>> 1.5 s
>>>> are from assigning the PCI devices to the 142 IOMMU groups.
>>>>
>>>> ```
>>>> $ lspci | wc -l
>>>> 281
>>>> $ dmesg
>>>> […]
>>>> [ 2.918411] PCI: Using host bridge windows from ACPI; if
>>>> necessary, use "pci=nocrs" and report a bug
>>>> [ 2.933841] ACPI: Enabled 5 GPEs in block 00 to 7F
>>>> [ 2.973739] ACPI: PCI Root Bridge [PC00] (domain 0000 [bus 00-16])
>>>> [ 2.980398] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM ClockPM Segments MSI HPX-Type3]
>>>> [ 2.989457] acpi PNP0A08:00: _OSC: platform does not support [LTR]
>>>> [ 2.995451] acpi PNP0A08:00: _OSC: OS now controls [PME
>>>> PCIeCapability]
>>>> [ 3.001394] acpi PNP0A08:00: FADT indicates ASPM is unsupported, using BIOS configuration
>>>> [ 3.010511] PCI host bridge to bus 0000:00
>>>> […]
>>>> [ 6.233508] system 00:05: [io 0x1000-0x10fe] has been reserved
>>>> [ 6.239420] system 00:05: Plug and Play ACPI device, IDs PNP0c02 (active)
>>>> [ 6.239906] pnp: PnP ACPI: found 6 devices
>>>
>>> For ~280 PCI devices, (6.24-2.92)/280 = 0.012 s/dev. On my laptop I
>>> have about (.66-.37)/36 = 0.008 s/dev (on v5.4), so about the same
>>> ballpark.
>>
>> Though if it was on average 0.008 s/dev here, around a second could be
>> saved.
>>
>> The integrated Matrox G200eW3 graphics controller (102b:0536) and the
>> two Broadcom NetXtreme BCM5720 2-port Gigabit Ethernet PCIe cards
>> (14e4:165f) take 150 ms to be initialized.
>>
>> [ 3.454409] pci 0000:03:00.0: [102b:0536] type 00 class 0x030000
>> [ 3.460411] pci 0000:03:00.0: reg 0x10: [mem 0x91000000-0x91ffffff pref]
>> [ 3.467403] pci 0000:03:00.0: reg 0x14: [mem 0x92808000-0x9280bfff]
>> [ 3.473402] pci 0000:03:00.0: reg 0x18: [mem 0x92000000-0x927fffff]
>> [ 3.479437] pci 0000:03:00.0: BAR 0: assigned to efifb
>>
>> The timestamp in each line differs by around 6 ms. Could printing the
>> messages to the console (VGA) hold this up (line 373 to line 911 makes
>> (6.24 s-2.92 s)/(538 lines) = (3.32 s)/(538 lines) = 6 ms)?
>>
>> [ 3.484480] pci 0000:02:00.0: PCI bridge to [bus 03]
>> [ 3.489401] pci 0000:02:00.0: bridge window [mem 0x92000000-0x928fffff]
>> [ 3.496398] pci 0000:02:00.0: bridge window [mem 0x91000000-0x91ffffff 64bit pref]
>> [ 3.504446] pci 0000:04:00.0: [14e4:165f] type 00 class 0x020000
>> [ 3.510415] pci 0000:04:00.0: reg 0x10: [mem 0x92e30000-0x92e3ffff 64bit pref]
>> [ 3.517408] pci 0000:04:00.0: reg 0x18: [mem 0x92e40000-0x92e4ffff 64bit pref]
>> [ 3.524407] pci 0000:04:00.0: reg 0x20: [mem 0x92e50000-0x92e5ffff 64bit pref]
>> [ 3.532402] pci 0000:04:00.0: reg 0x30: [mem 0xfffc0000-0xffffffff pref]
>> [ 3.538483] pci 0000:04:00.0: PME# supported from D0 D3hot D3cold
>> [ 3.544437] pci 0000:04:00.0: 4.000 Gb/s available PCIe bandwidth, limited by 5.0 GT/s PCIe x1 link at 0000:00:1c.5 (capable of 8.000 Gb/s with 5.0 GT/s PCIe x2 link)
>> [ 3.559493] pci 0000:04:00.1: [14e4:165f] type 00 class 0x020000
>>
>> Here is a 15 ms delay.
>>
>> [ 3.565415] pci 0000:04:00.1: reg 0x10: [mem 0x92e00000-0x92e0ffff 64bit pref]
>> [ 3.573407] pci 0000:04:00.1: reg 0x18: [mem 0x92e10000-0x92e1ffff 64bit pref]
>> [ 3.580407] pci 0000:04:00.1: reg 0x20: [mem 0x92e20000-0x92e2ffff 64bit pref]
>> [ 3.587402] pci 0000:04:00.1: reg 0x30: [mem 0xfffc0000-0xffffffff pref]
>> [ 3.594483] pci 0000:04:00.1: PME# supported from D0 D3hot D3cold
>> [ 3.600502] pci 0000:00:1c.5: PCI bridge to [bus 04]
>>
>> Can the 6 ms – also from your system – be explained by the PCI
>> specification? Seeing how fast PCI nowadays is, 6 ms sounds like a
>> long time. ;-)
>>
>>> Faster would always be better, of course. I assume this is not really
>>> a regression?
>>
>> Correct, as far as I know of, this is no regression.
>>
>>>> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device [8086:2034]
>>>> [ 6.996063] PCI: CLS 0 bytes, default 64
>>>> [ 7.000008] Trying to unpack rootfs image as initramfs...
>>>> [ 7.065281] Freeing initrd memory: 5136K
>>
>> The PCI resource assignment(?) also seems to take 670 ms:
>>
>> [ 6.319656] pci 0000:04:00.0: can't claim BAR 6 [mem 0xfffc0000-0xffffffff pref]: no compatible bridge window
>> […]
>> [ 6.989016] pci 0000:d7:05.0: disabled boot interrupts on device [8086:2034]
>>
>>>> […]
>>>> [ 7.079098] DMAR: dmar7: Using Queued invalidation
>>>> [ 7.083983] pci 0000:00:00.0: Adding to iommu group 0
>>>> […]
>>>> [ 8.537808] pci 0000:d7:17.1: Adding to iommu group 141
>>>
>>> I don't have this iommu stuff turned on and don't know what's
>>> happening here.
>>
>> There is a lock in `iommu_group_add_device()` in `drivers/iommu/iommu.c`:
>>
>> mutex_lock(&group->mutex);
>> list_add_tail(&device->list, &group->devices);
>> if (group->domain && !iommu_is_attach_deferred(group->domain, dev))
>> ret = __iommu_attach_device(group->domain, dev);
>> mutex_unlock(&group->mutex);
>>
>> No idea, if it’s related. Unfortunately, it’s a production system, so
>> I can’t do any debugging. (Maybe `initcall_debug` could give some
>> insight.) Maybe the IOMMU developers can explain it without it. Could
>> the IOMMU group assignment be done in parallel?
>
> FWIW I'd expect that locking to be pretty much immaterial - many devices
> are getting their own uncontended groups, and callers of this tend to be
> serialised at a higher level anyway. iommu_probe_device() usually runs
> off the back of the device_add() notifier (where it could be that it's
> the only thing making noise in between something *else* being slow), but
> there is the special case where it gets replayed for all existing
> devices when the IOMMU driver registers itself - at a guess it seems
> like it may well be the latter case you're seeing, but either way
> there's not much to say without figuring out where the time is actually
> being spent (I don't suppose that machine has dynamic ftrace enabled?).
Our Linux kernel has dynamic ftrace enabled.
$ grep CONFIG_DYNAMIC_FTRACE /boot/config-5.10.70.mx64.403
CONFIG_DYNAMIC_FTRACE=y
CONFIG_DYNAMIC_FTRACE_WITH_REGS=y
CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS=y
> That said, setting up a new group isn't a completely insignificant
> amount of work, and 142 groups seems a lot - I'd have assumed that a
> system of that scale would be the kind of big server kit that takes
> several minutes to boot to the point of even starting the kernel anyway.
You are right. As noted in my reply to Krzysztof, it’s more like a pet
peeve, but is also relevant, when kexec is used.
Kind regards,
Paul
Hi Paul,
> Thank you for your reply.
Thank you for getting back to us with a good insight.
[...]
> > I am curious - why is this a problem? Are you power-cycling your servers
> > so often to the point where the cumulative time spent in enumerating PCI
> > devices and adding them later to IOMMU groups is a problem?
> >
> > I am simply wondering why you decided to signal out the PCI enumeration as
> > slow in particular, especially given that a large server hardware tends to
> > have (most of the time, as per my experience) rather long initialisation
> > time either from being powered off or after being power cycled. I can take
> > a while before the actual operating system itself will start.
>
> It’s not a problem per se, and more a pet peeve of mine. Systems get faster
> and faster, and boottime slower and slower. On desktop systems, it’s much
> more important with firmware like coreboot taking less than one second to
> initialize the hardware and passing control to the payload/operating system.
> If we are lucky, we are going to have servers with FLOSS firmware.
>
> But, already now, using kexec to reboot a system, avoids the problems you
> pointed out on servers, and being able to reboot a system as quickly as
> possible, lowers the bar for people to reboot systems more often to, for
> example, so updates take effect.
A very good point about the kexec usage.
This is definitely often invaluable to get security updates out of the door
quickly, update kernel version, or when you want to switch operating system
quickly (a trick that companies like Equinix Metal use when offering their
baremetal as a service).
> > We talked about this briefly with Bjorn, and there might be an option to
> > perhaps add some caching, as we suspect that the culprit here is doing PCI
> > configuration space read for each device, which can be slow on some
> > platforms.
> >
> > However, we would need to profile this to get some quantitative data to see
> > whether doing anything would even be worthwhile. It would definitely help
> > us understand better where the bottlenecks really are and of what magnitude.
> >
> > I personally don't have access to such a large hardware like the one you
> > have access to, thus I was wondering whether you would have some time, and
> > be willing, to profile this for us on the hardware you have.
> >
> > Let me know what do you think?
>
> Sounds good. I’d be willing to help. Note, that I won’t have time before
> Wednesday next week though.
Not a problem! I am very grateful you are willing to devote some of you
time to help with this.
I only have access to a few systems such as some commodity hardware like
a desktop PC and notebooks, and some assorted SoCs. These are sadly not
even close to a proper server platforms, and trying to measure anything on
these does not really yield any useful data as the delays related to PCI
enumeration on startup are quite insignificant in comparison - there is
just not enough hardware there, so to speak.
I am really looking forward to the data you can gather for us and what
insight it might provide us with.
Thank you again!
Krzysztof
Dear Krzysztof,
Am 10.11.21 um 00:10 schrieb Krzysztof Wilczyński:
> [...]
>>> I am curious - why is this a problem? Are you power-cycling your servers
>>> so often to the point where the cumulative time spent in enumerating PCI
>>> devices and adding them later to IOMMU groups is a problem?
>>>
>>> I am simply wondering why you decided to signal out the PCI enumeration as
>>> slow in particular, especially given that a large server hardware tends to
>>> have (most of the time, as per my experience) rather long initialisation
>>> time either from being powered off or after being power cycled. I can take
>>> a while before the actual operating system itself will start.
>>
>> It’s not a problem per se, and more a pet peeve of mine. Systems get faster
>> and faster, and boottime slower and slower. On desktop systems, it’s much
>> more important with firmware like coreboot taking less than one second to
>> initialize the hardware and passing control to the payload/operating system.
>> If we are lucky, we are going to have servers with FLOSS firmware.
>>
>> But, already now, using kexec to reboot a system, avoids the problems you
>> pointed out on servers, and being able to reboot a system as quickly as
>> possible, lowers the bar for people to reboot systems more often to, for
>> example, so updates take effect.
>
> A very good point about the kexec usage.
>
> This is definitely often invaluable to get security updates out of the door
> quickly, update kernel version, or when you want to switch operating system
> quickly (a trick that companies like Equinix Metal use when offering their
> baremetal as a service).
>
>>> We talked about this briefly with Bjorn, and there might be an option to
>>> perhaps add some caching, as we suspect that the culprit here is doing PCI
>>> configuration space read for each device, which can be slow on some
>>> platforms.
>>>
>>> However, we would need to profile this to get some quantitative data to see
>>> whether doing anything would even be worthwhile. It would definitely help
>>> us understand better where the bottlenecks really are and of what magnitude.
>>>
>>> I personally don't have access to such a large hardware like the one you
>>> have access to, thus I was wondering whether you would have some time, and
>>> be willing, to profile this for us on the hardware you have.
>>>
>>> Let me know what do you think?
>>
>> Sounds good. I’d be willing to help. Note, that I won’t have time before
>> Wednesday next week though.
>
> Not a problem! I am very grateful you are willing to devote some of you
> time to help with this.
>
> I only have access to a few systems such as some commodity hardware like
> a desktop PC and notebooks, and some assorted SoCs. These are sadly not
> even close to a proper server platforms, and trying to measure anything on
> these does not really yield any useful data as the delays related to PCI
> enumeration on startup are quite insignificant in comparison - there is
> just not enough hardware there, so to speak.
>
> I am really looking forward to the data you can gather for us and what
> insight it might provide us with.
So, kexec seems to work besides some DMAR-IR warnings [1].
`initcall_debug` increases the Linux boot time by over 50 % from 7.7 s
to 12 s, which I didn’t expect.
Here are the functions taking more than 200 ms:
initcall pci_apply_final_quirks+0x0/0x132 returned 0 after 228433 usecs
initcall raid6_select_algo+0x0/0x2d6 returned 0 after 383789 usecs
initcall pcibios_assign_resources+0x0/0xc0 returned 0 after 610757
usecs
initcall _mpt3sas_init+0x0/0x1c0 returned 0 after 721257 usecs
initcall ahci_pci_driver_init+0x0/0x1a returned 0 after 945094 usecs
initcall pci_iommu_init+0x0/0x3f returned 0 after 1487134 usecs
initcall acpi_init+0x0/0x349 returned 0 after 7291015 usecs
Some of them are run later though, but `acpi_init` sticks out with 7.3 s.
Kind regards,
Paul
[1]:
https://lore.kernel.org/linux-iommu/[email protected]/T/#u