Linux Kernel Podcast for 2017/03/28


Author’s Note: Apologies to Ulrich Drepper for incorrectly attributing his paper “Futexes are Tricky” to Rusty. Oops. In any case, everyone should probably read Uli’s paper:

In this week’s edition: Linus Torvalds announces Linux 4.11-rc4, early debug with USB3 earlycon, upcoming support for USB-C in 4.12, and ongoing development including various work on boot time speed ups, logging, futexes, and IOMMUs.

Linus Torvalds announced Linux 4.11-rc4, noting that “So last week, I said that I was hoping that rc3 was the point where we’d start to shrink the rc’s, and yes, rc4 is smaller than rc3. By a tiny tiny sidgen. It does touch a few more files, but it has a couple fewer commits, and fewer lines changed overall. But on the whole the two are almost identical in size. Which isn’t actually all that bad, considering that rc4 has both a networking merge and the usual driver suspects from Greg [Kroah Hartman], _and_ some drm fixes”.


Junio C Hamano announced Git v2.12.2.

Greg Kroah-Hartman announced Linux 4.4.57, 4.9.18, and 4.10.6.

Sebastian Andrezej Siewior announced Linux v4.9.18-rt14, which includes a “larger rework of the futex / rtmutex code. In v4.8-rt1 we added a workaround so we don’t de-boost too early in the unlock path. A small window remained in which the locking thread could de-boost the unlocking thread. This rework by Peter Zijlstra fixes the issue.”

Upcoming features

Greg K-H finally accepted the latest “USB Type-C Connector class” patch series from Heikki Krogerus (Intel). This patch series aims to provide various control over the capability for USB-C to be used both as a power source and as a delivery interface to supply to power to external devices (enabling the oft-cited use case of selecting between charging your cellphone/mobile device or using said device to charge your laptop). This will land a new generic management framework exposed to userspace in Linux 4.12, including a driver for “Intel Whiskey Cove PMIC [Power Management IC] USB Type-C PHY”. Your author looks forward to playing. Greg thanked Heikki for the 18(!) iterations this patch went through prior to being merged – not quite a record, but a lot of effort!

Kishon Vijay Abraham (TI) posted “PCI: Support for configurable PCI endpoint”, which provides generic infrastructure to handle PCI endpoint devices (Linux operating as a PCI endpoint “device”), such as those based upon IP blocks from DesignWare (DW). He’s only tested the design on his “dra7xx” boards and requires “the help of others to test the platforms they have access to”. The driver adds a configfs interface including an entry to which userspace should write “start” to bring up an endpoint device. He adds himself as the maintainer for this new kernel feature.

Rob Herring posted “dtc updates for 4.12”, which “syncs dtc [Device Tree Compiler] with current mainline [dtc]”. His “primary motivation is to pull in the new checks [he’s] worked on. This gives lots of new warnings which are turned off by default”.

60Hz vs 59.94Hz (Handling of reduced FPS in V4L2)

Jose Abreu (Synopsys) posted a patch series entitled “Handling of reduced FPS in V4L2”, which aims to provide a mechanism for the kernel to measure (in a generic way) the actual Frames Per Second for a Video For Linux (V4L) video device. The patches rely upon hardware drivers being able to signal that they can distinguish “between regular fps and 1000/1001 fps”.

This took your author on a journey of discovery. It turns out that (most of the time), when a video device claims to be “60fps” it’s actually running at 59.94fps, but not always. The latter frame rate is an artifact of the NTSC (National Television System Committee) color television standard in the United States. Early televisions used the 60Hz frequency (which is nationally synchronized, at least in each of the traditional three independent grids operated in the US, which are now interconnected using HVDC interconnects but presumably are still not directly in phase with one another – feel free to educate me!) of the AC supply to lock individual frame scan times. When color TV was introduced, a small frequency offset was used to make room in each frame for a color sub-carrier signal while retaining backward compatibility for black and white transmissions. This is where frequencies of 29.97 and 59.95 frames per second originate. In case you always wondered.

Jose and Hans Verkuil had a back and forth discussion about various real- world measured pixelclock frequencies that they had obtained using a variety of equipment (signal analyzers, certified HDMI analyzer, and the Synopsys IP supported by the patch series under discussion) to see whether it was in reality possible to reliably distinguish frame rates.

Early Debug with USB3 earlycon (early printk)

Lu Baolu (Intel) posted version 8 of a patch series entitled “usb: early: add support for early printk through USB3 debug port”. Contemporary (especially x86) desktop and server class systems don’t expose low level hardware debug interfaces, such as JTAG debug chains, which are used during chip bringup and early firmware and OS enablement activities, and which allow developers with suitable tools to directly control and interrogate hardware state. Or just dump out the kernel ringbuffer (the dmesg “log”).

Actually, all such systems do have low level debug capabilities, they’re just fused out during the production process (by blowing efuses embedded into the processor) and either not exposed on the external pins of the chip at all, or are simply disabled in the chip logic. Probably most of these can be re-enabled by writing the magic cryptographically signed hashes to undocumented memory regions in on-chip coprocessor spaces. In any case, vendors such as Intel aren’t going to tell you how.

Yet it is often desirable to have certain low level debug functionality for systems that are deployed into field settings, even to reliably dump out the kernel console log DEBUG log level messages somewhere. Traditionally this was done using PC serial ports, but most desktop (and all laptop) systems no longer ship with those exposed on the rear panel. If you’re lucky you’ll see an IDC10 connector on your motherboard to which you can attach a DB9 breakout cable. Consumers and end users have no idea what any of this means, and in the case that they don’t know what this means, they probably shouldn’t be encouraged to open the machine up and poke things. Yet even in the case that IDC10 connectors exist and can be hooked up, this is still a cumbersome interface that cannot be relied upon today.

Microsoft (who are often criticized but actually are full of many good ideas and usually help to drive industry standardization for the broader market) instituted sanity years ago by working with the USB Implementors Forum (IF) to ensure that the USB3 specification included a standardized feature known as xHCI debug capability (DbC), an “optional but standalone functionality by an xHCI hosst controller”. This suited Windows, which traditionally requires two UARTs (serial ports) for kernel development, and uses one of them for simple direct control of the running kernel without going through complex driver frameworks. Debug port (which also existed on USB2) traditionally required a special external partner hardware dongle but is cleaner in USB3, requiring only a USB A-to-A crossover cable connecting USB3.0 data lines.

As Lu Baolu notes in his patch, “With DbC hardware initialized, the system will present a debug device through the USB3 debug port (normally the first USB3 port)”. The patch series enables this as a high speed console log target on Linux, but it could be used for much more interesting purposes via KDB.

[Separately, but only really related to console drivers and not debugging, Thierry Escande posted “firmware: google memconsole” which adds support for importing the boot time BIOS memory based console into the kernel ringbuffer on Google Coreboot systems].

Ongoing Development

Pavel Tatashin (Oracle) posted “parallelized “struct page” zeroing”, which improves boot time performance significantly in the case that the “deferred struct page initialization feature is enabled”. In this case, zeroing out of the kernel’s vmemmap (Virtual Memory Map) is delayed until after the secondary CPU cores on a machine have been started. When this is done, those cores can be used to run zeroing threads that write to memory, taking one SPARC system down from 97.89 seconds to boot down to 46.91. Pavel notes that the savings are also considerable on x86 systems too.

Thomas Gleixner had a lengthy back and forth with Pasha Tatashin (Oracle) over the latter’s posting of “Early boot time stamps for x86” which use the TSC (Time Stamp Counter) on Intel x86 Architecture. The goal is to log how long the machine actually took to boot, including firmware, rather than just how long Linux took to boot from the time it was started. Peter Zijlstra responded (to Pasha), “Lol, how cute. You assume TSC starts at 0 on reset” (alluding to the fact that firmware often does crazy things playing with the TSC offset or directly writing to it). Thomas was unimpressed with Pavel’s posting of a v2 patch series, noting “Did you actually read my last reply on V1 of this? I made it clear that the way this is done, i.e. hacking it into the earliest boo[]t stage is not going to happen…I don’t care about you wasting your time, but I very much care about my time”. He provided a further more lengthy response, including various commentary on the best ways to handle feedback.

Peter Zijlstra posted version 6 of a patch series entitled “The arduous story of FUTEX_UNLOCK_PI” in which he adds “Another installment of the futex patches that give you nightmares”. Futexes (Fast User-space Mutexes) are a mechanism provided by the Linux kernel which leverage shared memory to provide a low overhead mutex (mutual exclusion primitave) to userspace in the case that such mutexes are uncontended (no conflicts between processes – tasks within the kernel – exist trying to acquire the same resource) but with a “slow path” through the kernel in the case of contention. They are used by many userspace applications, including extensively in the C library (see the famous paper by Rusty Russell entitled “Futexes are Tricky”). Peter is working on solving problems introduced by having to have Priority Inheritance (PI) aware futexes in Real Time kernels. These adjust priority of the associated tasks holding mutexes for short periods in order to prevent Priority Inversion (see Mars Pathfinder study papers) in which a low priority task holds a mutex that a high priority task wants to acquire. Peter’s patches “rework[] and document[] the locking” of existing code.

Separately, Waiman Long (Red Hat) posted version 6 of “futex” Introducing throughput-optimized (TP) futexes which “introduces a new futex implementation called throughput-optmized (TP) futexes. It is similar to PI futexes in its calling convention, but provides better throughput than the wait-wake (WW) futexes by encouraging lock stealing and optimistic spinning. The new TP futexes an be used in implementing both userspace mutexes and rwlocks. The provide[] better performance while simplifying the userspace locking implementation at the same time. The WW futexes are still needed to implement other synchronization primitives like conditional variables and semaphores that cannot be handled by the TP futexes”.

David Woodhouse posted “PCI resource mmap cleanup” which aims to clean up the use of various kernel interfaces that provide “user visible” resource addresses through (legacy) proc and (contemporary) sysfs. The purpose of these interfaces is to provide information about regions of PCI address space memory that can be directly mapped by userspace applications such as those linked against the DPDK (Data Plane Development Kit) library. An example of his cleanup included “Only allow WC [Write Combining] mmap on prefetchable resources” for the /proc/bus/pci mmap interface because this was the case for the preferred sysfs interface already. This lead some to debate why the 64-bit ARM Architecture didn’t provide the legacy procfs interface (since there was a little confusion about the dependencies for DPDK) but ultimately re-concluded that it shouldn’t.

Tyler Baicar (Codeaurora) posted version 13 of a patch series entitled “Add UEFI 2.6 and ACPI 6.1 updates for RAS on ARM64”, which aims to introduce support to the 64-bit ARM Architecture for logging of RAS events using the shared “GHES” (Generic Hardware Error Source) memory location “with the proper GHES structures to notify the OS of the error”. This dovetails nicely with platforms performing “firmware first” error handling in which errors are trapped to secure firmare which first handles them and subsequently informs the Operating System using this ACPI feature.

Shaohua Li (Facebook) posted a patch entitled “add an option to disable iommu force on” in the case of the (x86) Trusted Boot (TBOOT) feature being enabled. The reason cited was that under a certain 40GBit networking load XDP (eXpress Data Path) test there were high numbers of IOTLB (IO Translation Look Aside Buffer) misses “which kills the performance”. What he is refering to is the mechanism through which an IOMMU (which sits logically between a hardware device, such as a network card, and memory, often as part of an integrated PCI Root Complex) translates underlying memory accesses by the adapter card into real host memory transactions. These are cached by the IOMMU in small caches (known as IOTLBS) after it performs such translations using its “page tables” (similar to how a host CPU’s MMU – Memory Management Unit – performs host memory translations). Badly designed IOMMU implementations or poor utilization can result in large numbers of misses that result in users disabling the feature. Alas, without an IOMMU, there’s little protection during boot from rogue devices that maliciously want to trash host memory. Nobody has noted this in the RFC (Request For Comments) discussion, yet.

Bodong Wang (Mellanox) posted a patch entitled “Add an option to probe VFs or not before enabling SR-IOV”, which aims to allow administrators to limit the probing of (PCIe) Virtual Functions (VFs) on adapters that will have those resources passed through to Virtual Machines (VMs) (using VFIO). This “can save host side resource usage by VF instances which would be eventually probed to VMs”. It adds a new sysfs interface to control this.

Viresh Kumar posted a patch entitled “cpufreq: Restore policy min/max limits on CPU online”. Apparently, existing code behavior was that “On CPU online the cpufreq core restores the previous governor [the in kernel logic that determines CPU frequency transitions based upon various metrics, such as saving energy, or prioritizing performance]…but it does not restore min/max limits at the same time”. The patch addresses this shortcoming.

Wanpeng Li posted a patch entitled “KVM: nVMX: Fix nested VPID vmx exec control” that aims to “hide and forbid” Virtual Processor IDentifiers in nested virtualization contexts where the hardware doesn’t support this. Apparently it was unconditionally being enabled (based upon real hardware realities of existing implementation) regardless of feature information (INVVPID) provided in the “vmx” capabilities.

Joerg Roedel posted a patch entitled “ACPI: Don’t create a platform_device for IOAPIC/IOxAPIC” since this was causing problems during hot remove (of CPUs). Rafael J. Wysocki noted that “it’s better to avoid using platform_device for hot-removable stuff” since it is “inherently fragile”.

Kees Cook (Google) posted a patch disabling hibernation support on 32-bit systems in the case that KASLR (Kernel Address Space Layout Randomization) was enabled at boot time, but allowing for “nokaslr” on the kernel command line to change this. Evgenii Shatokhin initially noted that “nokaslr” didn’t re-enable hibernation support correctly, but eventually determined that the ordering and placement of the “nokaslr” on the command line was to blame, which lead to Kees saying he would look into the command line parsing sequence and interaction with other options, such as “resume=”.

Separately, Baoquan He (Red Hat) noted that with KASLR an implicit assumption that EFI_VA_START < EFI_VA_END existed, while “In fact [the] EFI [(Unified) Extensible Firmware Interface] region reserved for runtime services [these are callbacks into firmware from Linux] virtual mapping will be allocated using a top-down schema”. His patches addressed this problem, and being “RESEND”s, he was keen to see that they get taken up soon.

Also separately, Kees posted “syscalls: Restore address limit after a syscall” which “ensures a syscall does not return to user-mode with a kernel address limit. If that happened, a process can corrupt kernel-mode memory and elevate privileges”. He cites a bug it would have prevented.

Kan Liang (Intel) posted “measure SMI cost”. This patch series aims to leverage hardware counters to inform perf of the amount of time spent (on Intel x86 Architecture systems) inside System Management Mode (SMM). SMIs (System Management Interrups) are events that are generated (usually) by Intel Platform Control Hub and similar chipset logic which can be programmed by firmare to generate regular interrupts that target a secure execution context known as SMM (System Management Mode). It is here that firmware temporarily steals CPU cycles from the Operating System (without its knowledge) to perform such things as CPU fan control, errata handling, and wholesale VGA graphics emulation in BMC “value add” from OEMs). Over the years, the amount of gunk hidden in SMIs has grown that this author even once wrote a latency detector (hwlat) and has a patent on SMI detection without using such dedicated counters…due to the impact of such on system performance. SMM is necessary on x86 due to its lack of a standardized on-SoC platform management controller, but so is accounting for bloat.

Finally, yes, Kirill A. Shutemov snuck in another iteration of his Intel “5-level paging support” in preparation for a 4.12 merge.


Leave a Reply

Your email address will not be published. Required fields are marked *