Kernel Podcast for March 13th, 2017


In this week’s kernel podcast: Linus Torvalds announces Linux 4.11-rc2 (including pre-enablement for Intel 5-level paging), VMA based swap readahead, and ongoing development ahead of the next cycle.

Linus Torvalds announced Linux 4.11-rc2. In his announcement, he said that the past week had been “fairly quiet” because “people are still looking for bugs and taking a breather after the merge window”. But he also noted that “we’ve got a healthy number of fixes in, and there’ssome cleanup/prep patches for the upcoming 5-level page table support that I took after the merge window just to make the next merge window easier”.

Various fixes and updates have been posted against the previous rc1, over the past week, including an urgent fix from Matthew (Willy) Wilcox for his idr rewrite in 4.11 (freeing the correct IDA bitmap).

Geert Uytterhoeven posted “Build regressions/improvements in v4.11-rc1”. This compared build error/warning regressions and improvements between v4.11-rc1 and v4.10. According to Geert, the 4.11-rc1 kernel saw an increase of 19 build errors and 1108 warnings when compared to 4.10.


Jiri Slaby announced Linux 3.12.71, Greg Kroah Hartman (KH) announced 4.4.53, 4.9.14, and 4.10.2 (which started a conversation about git tags being stale that we will address in a moment). Greg took the opportunity of various stable kernel work to prod the i915 graphics driver team with a message entitled “The i915 stable patch marking is totally broken”.

Sebastian Andrzej Siewior announced the v4.9.13-rt12 preempt-rt “Real Time” kernel patch set, which has a known issue that “CPU hotplug got a little better but can deadlock”, suggesting you might not want to try that then.

Julia Cartwright announced 4.1.38-rt46.

Steven Rostedt announced the 3.18.48-rt53 stable release of the RT kernel. He also announced the 3.10.105-rt119 and 3.2.86-rt124 releases.

Jair Ruusu announced “loop-AES-v3.7k file/swap crypto package”, which is available on sourceforge at:

Andy Lutomirski sent out detailed notes (along with a followup with yet more explanation) of the Intel SGX (“Secure Enclave”) feature discussion that occured at Kernel Summit and Linux Plumbers Conference last fall. The thread is called “SGX notes from KS/LPC”. In the thread, he explains what SGX is (a small region of virtual memory within a Linux process – known as a task inside the kernel – that is not visible to the host OS after the enclave is “launched”) and how it can be used to hide certain data from system administrators or providers – for example, cryptographic keys that one would rather were not compromised. SGX comes with a litany of new requirements upon the Operating System that Andy covers, along with some guidelines for how to expose this feature, and how to make it useable. are now sponsoring the project to the tune of various geo-diverse bare metal frontend systems in datacenters around the globe. Each of these (powerful) frontends provides read-only public access to git repositories and the public website ( and More information, including machine specifications can be found here:

(this came to light because of a brief outage affecting the Newark, NJ mirror which was lagging behind other mirrors in picking up new git tags pushed, but one hopes that an official announcement and thanks was otherwise forthcoming)

Masahiro Yamada has been added as a Kbuild (co-)maintainer.

Intel 5-level paging

Kirill A. Shutemov posted version 4 of his “5-level paging” patch series that implements support for the la57 (56 bit Virtual Address space for x64 Canonical Addressing) feature on some future CPUs. We covered the underlying patch series before, explaining the benefit of a larger (virtual) address space, as well as the additional compexities required to implement backward compatibility (including new prctls to limit the virtual address space of certain legacy applications), and the lack (so far) of boot time switching between 4-and-5-level support, which is seen as important for the distros.

Linus responded by saying that he thought “we should just aim for this being in 4.12” as he didn’t “see any real reason to delay merging it”. After some discussion about whose tree to merge it through, it was decided (by Thomas Gleixner) that it could come in through the “-tip” x86 tree. Which resulted in Linus pulling a preparatory “5-level paging: prepare generic code” patch series from Kirill into 4.11 (even after the merge window had closed) to lay the groundwork for pulling the main feature into the next (4.12) cycle. This promptly broke PowerPC, which was promptly fixed by a followup patch. Following the merge of enabling support in 4.11, Kirill posted “5-level paging enabling for v4.12” which aims to complete the merge next cycle.

The earlier version 4 iteration of the patch series noted that the Xen hypervisor currently doesn’t support 5-level paging and thus CONFIG_XEN is disabled automatically when building CONFIG_X86_5LEVEL. It was pointed out by the Andrew Cooper that runtime (boottime) switching between 4 and 5 level support would be required in order to provide a clean experience, especially until Xen Dom0 support is available. That boottime switching is on the existing todo and presumably is going to land at some point.

Separately, Dmitry Safonov posted version 6 of a patch series entitled “Fix compatible mmap() return pointer over 4Gb” which has “some minor conflicts with Kirill’s set for 5-table paging”. Dmitry aims to solve a slightly different problem than Kirill’s PR_{SET,GET}_MAX_VADDR calls (which limit the virtual address ranges returned by mmap to avoid legacy programs breaking when suddenly able to receive much larger “Canonical Addresses” – in Intel parlance – than they were compiled with built-in and broken assumptions about once upon a time) insomuch as he is focused on 32-bit legacy syscalls on 64-bit x64 not returning memory above 4GB that cannot be used by older 32-bit code.

VMA based swap readahead

Ying Huang (Intel) posted an RFC (Request For Comments) entitled “mm, swap: VMA based swap readahead” in which he discussed the current kernel paging implementation for Virtual Memory Areas (VMAs) as well as how it could be improved to facilitate greater awareness of the in-memory access patterns of associated data by changing the corresponding readahead algorithm.

“Readahead” as a concept is what it sounds like. Locality (both spacial, in this case, as well as temporal, in other cases) of data means that when a memory access occurs, it is usually more likely than not that an access to a nearby memory location will soon follow (except in the case of pure random access workloads). Thus, the kernel contains support for preloading nearby data when performing various disk and memory operations. Examples include readahead of nearby disk blocks when loading filesystem data, and loading nearby disk blocks when reading pages back in from swap.

VMAs (Virtual Memory Areas) are regions of memory managed by the Linux kernel. A running application (process), known as a “task” by the kernel, contains a large number of different VMAs which form its overall address space. You can see this by inspecting /proc/self/maps (replacing “self” with a process ID that you have access to). The output will show a series of memory regions representing various memory owned by the task. Memory that doesn’t represent files is known as “anonymous memory” and it is what is paged (swapped) out under memory pressure situations.

As Ying notes in his RFC, the “original swap readahead algorithm does readahead based on the consecutive blocks in [the] swap device” but “the consecutive blocks in [the] swap device just reflect the order of page reclaiming” and not necessarily “the access sequence in RAM”. His patch series aims to change this by teaching the readahead algorithm about VMAs and how to bias the readahead to sequentially walk through the address space of a task (process), reading those parts of the swap space containing this data rather than simply walking through swap sequentially.

But wait! There’s more! Ying also posted a separate patch series entitled “THP swap: Delay splitting THP during swapping out”, which does what it sounds like it would do. THP (Transparent Huge Pages) is a technology used by the Linux kernel to dynamically allocate “huge” (optionally very large – up to 1GB in size, but in this case 2MB) pages of memory to contiguous regions of virtual memory address space, especially those backing shared large memory data (even including a huge zero page used for virtual machine RAM at boot). THP reduces pressure on limited CPU internal microarchitectural caches known as TLBs (Translation Lookaside Buffers) – as well as uTLBs at a lower level than the TLBs – which cache the translation performed by page table entries to physical or intermediate memory addresses. Reducing the number of TLBs required to map regions of virtual memory reduces the number of times TLBs must be reused by the underlying architecture during memory access operations.

The existing Linux kernel THP code splits THPs back into smaller pages whenever they are swapped (paged) out to disk. Yet it turns out that this is particularly inefficient on contemporary systems in which secondary disk or NVMe storage has far greater bandwidth than a single high end core can saturate if forced to do this work. Ying’s patch instead delays this split and pushes entire THPs out to swap, allowing for larger writes and reads of contiguous memory out to the backing storage.

Ongoing Development

“David F” inquired about RAID mode support for Intel m.2 chipsets. These devices continue the recent-ish legacy of certain Intel storage devices providing dual modes of operation: as an AHCI device, and as a hardware RAID device operating in a propietary mode for which no Linux drivers exist. David was quite concerned that the lack of a Linux driver was becoming particular problematic on newer machines, which might not provide a means to switch into AHCI mode (supported by Linux). Christoph Hellwig was…unsympathetic…suggesting that the RAID mode “provides worse performance”, and that its implementation was questionable. He also had a series of other suggestions for what to do with these devices – those are less family friendly to repeat in this podcast.

Michal Hocko posted “kvmalloc” which is a generic replacement for the many “open coded kmalloc with vmalloc fallback instances in the tree”. k-and-vmalloc are two different means by which kernel code allocates memory. The former is used to obtain small allocations (on the order of a few pages – the minimal granule size operated on by the virtual memory subsystem of Linux on contemporary processors) that are also linerally contiguous in physical memory. The latter is for larger allocations of strictly “virtual” memory – contiguous only when accessed using the underlying Memory Mangement Unit to perform a translation (this is usually automatic for kernel code, since the kernel runs with virtual memory of its own, just like user processes do, but it can be problematic if a driver would like to use this memory for certain hardware operations, such as DMA transfers). The generic wrapper aims to clean up the common case that kernel code just wants a chunk of memory and will try to allocate it with kmalloc, but fallback to the more generic vmalloc if that fails.

Christian Konig (AMD) posted “PCI: add resizeable BAR infrastructure” (version 2, and later an update with some fixes in a version 3 also), which aims to add support to the kernel for a PCI SIG (Peripheral Component Interconnect Special Interest Group) ECN (Engineering Change Notice) that enables BARs (Base Address Registers) to be resized at runtime. PCI(e) BARs are mapping windows (aperatures) in the system memory map that are used to talk to hardware add-on cards (or built-in devices within modern platforms) by determining where the device’s memory will live. Traditionally, BARs were fixed size and so on architectures not relying upon firmware configuration of underlying BARs, Linux would have to determine where to place certain PCI(e) resources at boot/hotplug time by checking how much memory a device needed to expose and programming the BARs. With the new extension comes the possibility to increase the size of a BAR to map larger regions of memory. This is a useful feature for graphics cards, which may want to map very large regions of memory. A subsequent patch wires up the AMD GPU driver to use this.

Javi Merino posted “Documentation/EDID fixes”, which aims to correct some broken assumptions in the kernel documentation for EDID (Extended Display Identification Data – the data provided over e.g. I2C from a VGA monitor when the cable is connected). The examples didn’t build correctly due to existing assumptions. This author is probably one of few people who always thinks of EDID and the interaction with Xorg every time he plugs in an external projector to his laptop.

David Howells posted “net: Work around lockdep limitation in sockets that use sockets” in which he corrected an erroneous assumption in the kernel “lockdep” (lock dependency checker) that prevented it from correctly identifying bad call chains involving TCP sockets when there exists a dependency between sockets created purely in the kernel and sockets created purely in userspace (which the lockdep could not distinguish between due to its use of broad lock classes). The AFS (Andrew File System) was generating a false lockdep warning because it was exposing such an implied dependency.

Charles Keepax posted “genirq: Add support for nested shared IRQs” to address an audio CODEC that also acts as an interrupt controller. The details sounded rather painful. Yet it was “fairly easy” to fix.

Steven Rostedt posted “tracing: Allow function tracing to start earlier in boot up”, which does roughly what it says on the can, “moving tracing up further in the boot process”, “right after memory is initialized”. He noted that his RFC was a start and could be futher improved upon.

Matthew (Willy) Wilcox posted an RFC entitled “memset_l and memfill” that provides a generic means for architectures to provide optimized functions that “fill regions of memory with patterns larger than those contained in a single byte”. This is intended to be used by zram as well as other code.

Paul McKenney noticed some of his RCU torture tests failing during hotplug early in boot due to calls to smp_store_cpu_info during that operation. The call is not safe because it indirectly invokes schedule_work() which wants to use RCU prior to RCU being enabled as a side effect of dealing with an unstable TSC (Time Stamp Counter) on the afflicted CPU. Peter Zijlstra had an opinion on hotplug, and also a patch to handle this situation.

Vlad Zakharov posted “update timer frequencies”, which inquired about the best means to implement a cpufreq driver for ARC CPUs. These having a special property that “ARC timers (including those are used for timekeeping) are driven by the same clock as ARC CPU core(s)”. Yup, they change frequency according to the current CPU frequency. Which as Thomas Gleixner noted in response is “broken by design and you really should go and tell your hardware folks to fix that”. He added that “It’s well known for more than TWO decades that changing the frequency of the timekeeper clocksource is a complete disaster”.

Thomas Gleixner posted “kexec, x86/purgatory: Cleanup the unholy mess”, which aims to address his opinion that “the whole machinery is undocumented and lacks any form of forward declarations” (of variables which were previously global but had been made static). Purgatory is a special piece of code which is provided by the kernel but runs in the interim period between the kernel crashing (or beginning kexec) and the new crash or kexec kernel that is then subsequently loaded – this is what performs the load and exec.

Leave a Reply

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