diff options
Diffstat (limited to 'presentation/presentation.tex')
| -rw-r--r-- | presentation/presentation.tex | 229 |
1 files changed, 229 insertions, 0 deletions
diff --git a/presentation/presentation.tex b/presentation/presentation.tex new file mode 100644 index 0000000..b86129c --- /dev/null +++ b/presentation/presentation.tex @@ -0,0 +1,229 @@ +%------------------------------------------------------------------------------- +% presentation.tex - presentation of the QCOW2 in-kernel project +%------------------------------------------------------------------------------- +% author : Manuel Bentele +% date : Aug 26, 2019 +% copyright: (c) 2019 Manuel Bentele +%------------------------------------------------------------------------------- + +\documentclass[english]{beamer} + +%------------------------------------------------------------------------------- +% Presentation theme settings +%------------------------------------------------------------------------------- +\usetheme{Malmoe} +\usecolortheme{beaver} +\useinnertheme{rounded} + +%------------------------------------------------------------------------------- +% Used packages +%------------------------------------------------------------------------------- +\usepackage{colortbl} +\usepackage{xcolor} +\usepackage{hyperref} +\usepackage{tikz} +\usetikzlibrary{positioning, calc, arrows} +\usepackage{pgfplots} +\usepackage{pgfplotstable} +\pgfplotsset{compat = newest} +\usepackage{pifont} + +%------------------------------------------------------------------------------- +% Custom macros and definitions +%------------------------------------------------------------------------------- +\makeatletter\let\frametextheight\beamer@frametextheight\makeatother + +\newcommand{\cmark}{\ding{51}} +\newcommand{\xmark}{\ding{55}} + +\colorlet{hdr}{green!15} +\colorlet{ret}{yellow!15} +\colorlet{reb}{orange!15} +\colorlet{l1t}{red!15} +\colorlet{l2t}{violet!15} +\colorlet{dat}{cyan!15} + +%------------------------------------------------------------------------------- +% General definitions +%------------------------------------------------------------------------------- +\title{QCOW2 in the Linux kernel} +\author[M. Bentele]{Manuel Bentele} +\institute{University of Freiburg} +\date{September 2, 2019} + +\begin{document} + +\begin{frame} + \maketitle +\end{frame} + +\section{Analysis} +\subsection{Requirements} +\begin{frame}{What has to be done?} + \begin{block}{Implement the QCOW (QEMU Copy On Write) disk file format version~2 in the Linux kernel such that \dots} + \vspace*{0.5cm} + \begin{itemize} + \item reading of the normal QCOW2 disk file format is possible + \item compressed \& sparse QCOW2 disk files are supported as well + \item the disk file format is exposed as block device + \item the implementation compiles \& runs under Linux kernel 5 later + \item the performance is better than using qemu-nbd + \end{itemize} + \end{block} +\end{frame} + +\subsection{Linux storage stack} +\begin{frame}{How does the Linux storage stack looks like?} + \begin{center} + \resizebox{!}{0.775\textheight}{ + \input{images/linux_storage_stack} + } + \end{center} +\end{frame} + +\subsection{Implementation approaches} +\begin{frame}{How can the implementation be achieved?} + \begin{block}{\xmark\ FUSE (Filesystem in Userspace) driver} + \begin{itemize} + \item implement reading of QCOW2 file format as user space driver + \end{itemize} + \end{block} + \begin{block}{\xmark\ Device mapper target} + \begin{itemize} + \item implement reading of QCOW2 file format as mapped target + \end{itemize} + \end{block} + \begin{block}{\xmark\ Custom block driver} + \begin{itemize} + \item create block driver \& configuration utility for reading QCOW2 + \end{itemize} + \end{block} + \begin{block}{\cmark\ Loop device module extension} + \begin{itemize} + \item extend the loop device module \& configuration utility by a file format subsystem to implement QCOW2 as additional file format driver + \end{itemize} + \end{block} +\end{frame} + +\section{Implementation} +\subsection{Loop device module extension} +\begin{frame}{How is the file format subsystem integrated?} + \resizebox{\textwidth}{!}{% + \input{images/file_fmt_subsystem_integration} + } + \vfill + \begin{itemize} + \item file format subsystem abstracts the direct access to backing files to implement various file formats + \item file formats are implemented as file format drivers + \item drivers are registered at the subsystem + \item subsystem supports (asynchronous) reads, (asynchronous) writes, flushes and virtual disk sizes + \end{itemize} +\end{frame} + +\begin{frame}[fragile, shrink=35]{How does a file format driver look like?} +\begin{verbatim} +#include "loop_file_fmt.h" + +static int drv_file_fmt_read(struct loop_file_fmt *lo_fmt, + struct request *rq) { + /* TODO: implement reading of file format */ + return -EIO; +} + +static struct loop_file_fmt_ops drv_file_fmt_ops = { + .read = drv_file_fmt_read +}; + +static struct loop_file_fmt_driver drv_file_fmt = { + .name = "DRV", + .file_fmt_type = LO_FILE_FMT_RAW, + .ops = &drv_file_fmt_ops, + .owner = THIS_MODULE +}; + +// register driver with loop_file_fmt_register_driver(&drv_file_fmt) +// unregister driver with loop_file_fmt_unregister_driver(&drv_file_fmt) +\end{verbatim} +\end{frame} + +\subsection{QCOW2 file format driver} +\begin{frame}{How is the QCOW2 disk file format structured?} + \begin{columns} + \begin{column}[c]{0.25\textwidth} + \resizebox{\textwidth}{!}{% + \input{images/qcow2_structure} + } + \end{column} + \begin{column}[c]{0.75\textwidth} + \begin{itemize} + \item data is saved in data clusters of equal size (512\,B -- 2\,MB) + \item header provides offsets to 1st level tables + \item two-level lookup of data clusters (L1 \& L2~tables) + \item two-level reference count for copy on write (Refcount \& Refcount block tables) + \item numbers are stored in big-endian order + \item data clusters can be compressed or encrypted + \item support of embedded snapshots by use of internal copy on write + \end{itemize} + \end{column} + \end{columns} +\end{frame} + +\begin{frame}{How does QCOW2 addresses data clusters?} + \begin{itemize} + \item QCOW2 header stores an offset in the file to the L1~table + \item L1~table stores offsets in the file to L2 tables + \item L2~tables stores offsets in the file to the data clusters + \end{itemize} + \begin{center} + \resizebox{!}{0.55\textheight}{% + \input{images/qcow2_addressing} + } + \end{center} +\end{frame} + +\begin{frame}{How does the QCOW2 driver read data?} + \begin{columns} + \begin{column}[c]{0.4\textwidth} + \resizebox{\textwidth}{!}{% + \input{images/qcow2_reading} + } + \end{column} + \begin{column}[c]{0.6\textwidth} + Given a Linux IO read request with size $S$ and block device offset $O$: + \begin{enumerate} + \item calculate cluster $C$ and position $P$ for $O$ using cached L1 \& L2~tables + \item decompress the data of $C$ if $C$ is compressed + \item read data from $P$ into IO read request until $S$ bytes or the end of $C$ is reached + \item repeat steps 1 -- 3 until IO read request is filled with $S$ bytes + \end{enumerate} + \end{column} + \end{columns} +\end{frame} + +\subsection{Performance of the driver} +\begin{frame}{How does the implementation perform compared to qemu-nbd?} + \resizebox{\textwidth}{!}{ + \input{images/qcow2_performance} + } +\end{frame} + +\section{Outlook} +\subsection{Further work} +\begin{frame}{What can be done in the future?} + \begin{block}{File format subsystem} + \begin{itemize} + \item implement other file formats, e.g. VDI, VMDK, \dots + \item extend the API to support snapshots \& encryption + \end{itemize} + \end{block} + \begin{block}{QCOW2 file format driver} + \begin{itemize} + \item implement write operations + \item implement encryption \& snapshot support + \item improve performance by hardware aligned cache allocation + \item add a QCOW2 L2 cache clean interval + \end{itemize} + \end{block} +\end{frame} + +\end{document} |