Skip to content

aclements/sv6

Repository files navigation

sv6 is a POSIX-like research operating system designed for multicore scalability based on xv6.

sv6 is not a production kernel. Think of it as a playground full of half-baked experiments, dead code, some really cool hacks, and a few great results.

Building and running sv6 in QEMU

TL;DR: make && make qemu

You'll need GCC version 4.7 or later and GNU make.

There are several variables at the top of the top-level Makefile you may want to override for your build environment. It is recommended you set them in config.mk.

The kernel is configured via param.h. If you're just running sv6 in QEMU, you don't have to modify param.h, but you may want to read through it.

The most important Makefile variable is HW. This controls the hardware target you're building for and affects many settings both in the Makefile and param.h. The default HW is qemu. Each of our multicore machines also has a HW target (like josmp and ben), and other interesting HW targets are mentioned below. Builds go to o.$HW.

Common problems

panic: unhandled inode 369 type 0 on boot: This seems to be a bug in the virtual IDE controller of some versions of QEMU (though we're not positive). Try upgrading (or downgrading) QEMU.

Running sv6 on real hardware

Make sure you can build and boot sv6 in QEMU first.

Start by adding a HW target to param.h using one of the "physical hardware targets" in param.h as a template.

For HW targets where MEMIDE is defined to 1 (the default), the file system image is baked directly into the kernel image. This makes it possible to boot a physical machine into the sv6 kernel with nothing but the kernel image itself, and without having to worry about messing up your disks.

The kernel image is o.$HW/kernel.elf. This file is multiboot-complaint, so both GRUB and SYSLINUX can boot it directly. You can also PXE boot this image over the network using PXELINUX (that's what we do).

Optional components

lwIP

To enable networking support, you'll need to clone lwIP. From the root of your sv6 clone,

git clone git://git.savannah.nongnu.org/lwip.git
(cd lwip && git checkout DEVEL-1_4_1 && patch -p1 < ../lwip.patch)
make clean

(If you are building another hardware target, be sure to set HW when invoking make clean.)

mtrace

sv6 can be run under an mtrace-enabled QEMU to monitor and analyze its memory access behavior. You'll need to build and install mtrace:

git clone https://github.com/aclements/mtrace.git

And build with HW=mtrace. If mtrace isn't cloned next to the sv6 repository, then set MTRACESRC in config.mk to the directory containing mtrace-magic.h.

To run under mtrace, make mtrace.out.

Supported hardware

Not much.

sv6 is known to run on five machines: QEMU, a 4 core Intel Core2, a 16 core AMD Opteron 8350, 48 core AMD Opteron 8431, and an 80 core Intel Xeon E7-8870. Given the range of these machines, we're optimistic about sv6's ability to run on other hardware. sv6 supports both xAPIC- and x2APIC-based architectures.

For networking, sv6 supports several models of the Intel E1000, including both PCI and PCI-E models. If you have an E1000, you'll probably have to add your specific model number to the table in kernel/e1000.cc, but you probably won't have to do anything else.

Running sv6 user-space in Linux

Much of the sv6 user-space can also be compiled for and run in Linux using make HW=linux. This will place Linux-compatible binaries in o.linux/bin.

You can also boot a Linux kernel into a pure sv6 user-space! make HW=linux also builds o.linux/initramfs, which is a Linux initramfs file system containing an sv6 init, sh, ls, and everything else. You can boot this on a real machine, or run a super-lightweight Linux VM in QEMU using

make HW=linux KERN=path/to/Linux/bzImage/or/vmlinuz qemu

How to

CPU profiling

sv6 supports NMI-based system-wide hardware performance counter profiling on both Intel and AMD CPUs. On recent Intel CPUs, it also supports PEBS precise event sampling and memory load latency profiling.

To profile a command, use the perf tool. E.g.,

perf mailbench -a all / 1

By default, perf monitors unhalted CPU cycles, but other events can be selected from those known to libutil/pmcdb.cc.

Once perf has run, the sampler data can be read from /dev/sampler. To transfer the file to your computer where it can be decoded, use the web server:

curl http://<hostname>/dev/sampler > sampler

Finally, to decode the sample file, use perf-report:

./o.$HW/tools/perf-report sampler o.$HW/kernel.elf

To get stack traces from a user binary, pass its unstripped ELF image (e.g., o.$HW/bin/ls.unstripped) as the last argument instead of the kernel image.

Kernel statistics

The kernel continually maintains a lot of internal statistics counters. To see the changes in these counters over a command, run, e.g.

monkstats mailbench -a all / 1