Differences between revisions 126 and 127
|Deletions are marked like this.||Additions are marked like this.|
|Line 193:||Line 193:|
|KASLR:: '''K'''ernel '''A'''ddress '''S'''pace '''L'''ayout '''R'''andomization. A technique to make exploits harder by placing various objects at random, rather than fixed, addresses.|
A glossary of various terms and acronyms related to the Linux kernel. If you know something, please create yourself an account (UserPreferences) and add a term in alphabetical order. Don't be afraid to improve other people's explanations, the goal is to have a high quality document for the readers of this site. If everybody who reads this page adds a term each week, this glossary should be complete within a few months...
- 2Q algorithm
- 8259 PIC
Application Binary Interface, the interface of passed structures between the user processes (and libraries) and the kernel. For compatibility, it is important that these remain as static as possible (i.e. making sure that variables and structure members have the same bytesize as before, and in the same ordering). Occasionally breakage is necessary, requiring re-compilation of the user-space sources (note that this does not affect source-compatibility; that is a separate issue).
Advanced Configuration and Power Interface, replacement for APM that has the advantage of allowing O/S control of power management facilities as well exporting the set of hardware currently present on the system.
- Anticipatory Scheduler
Generally, used for something which doesn't have the usual associated object. For example an anonymous address space is not interested in user address space (that is, no process context). Some common ones are :
- Anonymous page
A page of memory that is not associated with a file on a file system. This can come from expanding the process's data segment with brk(), shared memory segments, or mmap() with a MAP_ANON or MAP_PRIVATE flag. MAP_PRIVATE, although it maps in data from a file, is considered anonymous because any changes do not get written back to the file (any dirty pages have to be moved to swap if the page is freed from main memory).
- Anonymous buffer
The buffer cache contains buffers of data on their way to/from the disk. An anonymous buffer is not associated with a file. One example of this is data from a deleted file - it will not be written to any file, but is kept around until it is flushed.
- big lock
kernel_lock, which locks the entire kernel from entry (no other task may run in the kernel code). It is recursive per process and dropped automatically when a process gives up the CPU, then regained on wake-up, in contrast to other spinlocks.
- bit error
Used colloquially to mean a single bit error in some memory address. Often due to faulty memory (ECC memory can correct single bit errors). Often results in fake oopsen, with addresses like 0x0008000. Also seen are values some small offset from zero, plus a bit error, which is where the value passed a NULL check due to the bit error, and then the kernel tried to access a structure member by means of the pointer, leading to the offset.
- block bitmap
- bottom-half handler
A set of standard kernel threads that execute tasks on a queue that have been registered with that type of bottom-half handler for execution. The code is run on return to user space or at the end of a hardware interrupt. In 2.3.43 a more general solution with softirqs and tasklets was implemented. Sometimes abbreviated to "bh", which should not be confused with buffer head, which is also abbreviated to "bh".
- bounce buffer
Big-reader locks, used when there are many contending for read access to a resource, and very few contending for writes (thus the balance is towards very fast read locking, and very slow write locking).
- buddy allocator
The memory allocation scheme used in the kernel. A vector of lists of free pages is kept, ordered by the size of the chunk (in powers of two). When a chunk is allocated, it is removed from the relevant list. When a chunk is freed back to the free pages pool, it is placed in the relevant list, starting from the top. If it is physically contiguous with a present chunk, they are merged and placed in the list above (i.e. where the chunks are twice the size), and this operation percolates up the vector. As regions are merged whenever possible, this design helps to reduce memory fragmentation. FIXME
- buffer cache
The buffer cache is a hash table of buffers, indexed by device and block number. LRU lists are maintained for the buffers in the various states, with separate lists for buffers of different sizes. With 2.3's unification of the buffer and page caches, each buffer head points to part or all of a page structure, through which the buffer's actual contents are available. FIXME
- buffer head
A structure containing information on I/O for some page in real memory. A buffer can be locked during I/O, or in several other states depending on its usage or whether it is free. Each buffer is associated with one page, but every page may have several buffers (consider the floppy on x86, where the I/O blocksize is 512 bytes, but each page is commonly 4096 bytes).
- bus mastering
- byte sex
- cache affinity
Where the cache of a CPU represents the current memory set used by a task, there is said to be cache affinity with that task. A good thing if the task is regularly scheduled on that CPU. See processor affinity.
- cache coherency
On an SMP system, ensuring that the local memory cache of each CPU is consistent with respect to the values which may be stored in other CPUs' caches, avoiding coherency problems such as the "lost update". This is achieved by the hardware in concert with the operating system.
- cache line
A section of the hardware cache, around 32 bytes large. Kernel structures are often designed such that the commonly-accessed members all fit into one cache-line, which reduces cache pollution. Structures such as this are cache line aligned.
- cache ping-pong
A hardware phenomenon in an SMP system, where two tasks on different CPUs are both accessing the same physical memory in a cache line. This means as each task runs, when it changes the memory, it must invalidate the other CPU's relevant cache line (to ensure cache coherency). Then, when the task on the other CPU runs, it must reload the cache line (as it's set invalid), before changing it. Repeat ad jocularum. A bad thing (TM). A common reason for putting a lock on a different cache line than the data mutexed by the lock : then the "other" task can grab and drop the lock without having to necessarily invalidate the cache line on the first CPU. FIXME
- cache pollution
Where during execution of a task, another task is scheduled onto that CPU which disrupts useful lines of the current cache contents, which will be used soon. That is, cache pollution is a non-optimal situation where the other process would have been bettered scheduled on a different CPU or at a different time. The aim is to minimise the need to replace cache lines, obviously increasing efficiency.
- call gate
Class Based Queueing, a hierarchical packet fair queueing qdisc. CBQ Homepage
- chroot jail
x86 assembler instructions for disabling and enabling interrupts, respectively. There are CPU-local and global variants of these. Code running with interrupts disabled must be fast, for obvious reasons (this is called interrupt latency).
Eric Raymond's proposal for a replacement to the current kernel build system. See http://www.tuxedo.org/~esr/kbuild.
- cold cache
- completion ports
Where two tasks each want an exclusive resource. You may hear talk of, for example, spinlock contention, which is where one or more tasks is commonly busy-waiting for a spinlock to become unlocked, as it is being taken by other tasks.
- Context switch
Refers to the changes necessary in the CPU when the scheduler schedules a different process to run on the CPU. This involves invalidating the TLB, loading the registers with the saved values, etc. There is an associated cost with such a switch, so it is best to avoid un-necessary context switch when possible. Note that the division of kernel-mode and user-mode means a similar, but simpler, operation is necessary when a syscall moves into kernel mode. However this is not called a context switch, as the mode switch doesn't change the current process. See lazy TLB. One good of feature of Linux is its extremely low context and mode switch cost, compared to an operating system like Solaris.
- critical path
A vital code path which should be optimised for the common case. Critical paths are executed frequently and form the important trunk routes of various kernel operations. An example would be buffer head manipulation during file I/O.
- Device Mapper
A technology for presenting arbitrary groupings of underlying sectors on physical devices in a consistent logical fashion usable by higher level algorithms. Heavily used by kernel technologies such as LVM.
- dancing makefiles
The cache of dentry structures. Under UNIX an entry in a particular directory must be searched for linearly, so even if the disk block containing the directory entry list is in-core, there is an associated cost. The dcache stores recent results of these searches which in general speeds up these disk searches by a large factor. Recent 2.3 work uses the dentries to allow multiple mounting, union mount, and more.
- delayed write
- demand zero
- directory notification
- drop behind
In stream I/O conditions, data that has already been read and processed is not needed again. The VM ideally should recognise this and mark the used pages as un-needed, so they can be discarded first. This technique is called "drop behind".
- eager coalescing
- edge-triggered interrupt
The interrupt is triggered by the rising or falling edge of the interrupt line. This makes IRQ line sharing difficult, as an edge may occur whilst an ISR is running, and it could be easily missed; to allow sharing level-triggered interrupts are usually used.
- elevator algorithm
This algorithm, often used in disk accesses, keeps an ordered list of requests. When the current request on the disk (e.g. the disk block) has been satisfied, the next strictly greater request on the list is dealt with. When a new request arrives, it is inserted into the ordered list in position (e.g. if the new requested block number is less than the current handled request, it goes before it in the list). When reaching the end of the list, the elevator changes direction, and the situation is reversed.
Explicitly-Parallel Instruction set Computing, an instruction set architecture where every dependency for an instruction is encoded into the instruction itself. This has the potential to be faster as the compiler can encode the data dependencies in the instructions.
- exponential back-off
- extended attributes
Also known as multi-part or multi-stream files, files with extended attributes deviate from the principle of files being a simple single data stream. An example of extended attributes is the Macintosh's "resource fork", which is associated with a specific file (known as the "data fork").
- fair scheduler
A scheduler which ensures fairness between users, such that a user's process count and associated cost only impacts that user, rather than the whole system as currently. Rik van Riel and Borislav Deianov have both produced different patches to implement this.
- false sharing
On SMP caches, when two parts of single block are accessed, neither of which collide with the other, the cache coherency protocol may not be able to detect this, and mark the block as "shared" even when it isn't. This is known as false sharing.
The code path most commonly taken, often optimised heavily at the expense of less frequently-taken blocks of code. This is the reason you see so many gotos in core functions - it produces common-path code far more efficient than an optimising compiler can manage.
- fixed mmap
A user-space request for a mmap starting at a fixed virtual address. Generally not useful or guaranteed to work; a notable exception is overlayed mmaps, where a mmaped area has further mmaps of different types at fixed positions in the map.
GNOME's source documentation system (similar to javadoc). Available by CVS from gnome. Kernel driver interface descriptions, built from source using gdoc, are currently being written in 2.3.
In the kernel, often means "get a reference to". This may be as simple as incrementing a usage count, or it may imply attempting to retrieve an object from a cache of some sort, or allocating kernel memory. See put.
- group descriptor
Global System Interrupt. Mainly used in the context of ACPI. Stupid acronym
high memory, or memory that is not permanently mapped into kernel memory. Common on 32 bit x86 systems. See HighMemory.
High Precision Event Timer (HPET) is a replacement timer for the 8254 Programmable Interval Timer and the Real-time clock's (RTC) periodic interrupt function. HPET is a successor to pmtimer, and is far more efficient to read.
The HPET can produce periodic interrupts at a much higher resolution than the RTC and is often used to synchronize multimedia streams, providing smooth playback and reducing the need to use other timestamp calculations such as an x86 cpu's RDTSC instruction. HPET support in linux requires that the BIOS expose the HPET (via acpi).
Hierarchical Token Bucket, a qdisc based on TBF and CBQ. HTB Theory
IP Virtual Server, the kernel part of the LVS (Linux Virtual Server) project. IPVS redirects incoming client requests to one of several "real" servers, usually for the purpose of load balancing a service.
An incrementing counter representing system "uptime" in ticks - or the number of timer interrupts since boot. Ultimately the entire original concept of a jiffy will likely vanish as systems use timer events only when necessary and become "jiffyless".
Logical Block Addressing. A way to address IDE disks without Cylinder/Head/Sector (CHS) coordinates, using linear sector numbers from the start of the disk. Allows for the use of very large IDE disks.
- Linux Device Drivers, 3rd Edition
Linux Kernel Mailing List. The primary virtual watering hole (meeting ground) for kernel developers to share ideas and bounce opinions off one another during the course of the kernel development process. FAQ at http://www.tux.org/lkml/.
Logical Volume Management. A technology for providing an arbitrary logical view of underlying data storage in a fashion supporting resizing and restructuring of storage on the fly. Currently in version 2, originally written by Sistina (now Redhat).
a cross-reference tool that can be used to navigate the Linux kernel source code, available at lxr.linux.no.
Memory Management Unit, part of the CPU hardware that enforces memory boundaries, and throw page faults, upon which the OS builds its coherent protection. The MMU maps virtual memory to actual, where protections allow.
MUTual EXclusion locks. This locking primitive is simpler and semantically tighter than the others, and hence is easier to make faster, and to prove correct. Some constraints are; lock has one owner at a time, the locker, who must also be the unlocker. Read Documentation/mutex-design.txt for much more.
Message Signaled Interrupts. A PCI mode where the interrupt numbers are extended from 8 bits to 32. These also use the normal pci data lanes not some magic all over the chipset; which means that a device can basically have as many interrupts as it wants rather than 4 (1 in practice) for legacy PCI interrupts, and there are also no interrupt sharing issues, since there are just so many numbers for interrupts... For more, see http://en.wikipedia.org/wiki/Message_Signaled_Interrupts
NAPI ("New API") is a modification to the device driver packet processing framework, which is designed to improve the performance of high-speed networking. See http://www.linux-foundation.org/en/Net:NAPI.
Netfilter is a framework that provides a set of hooks within the Linux kernel for intercepting and manipulating network packets. See http://en.wikipedia.org/wiki/Netfilter and http://www.netfilter.org.
- Page cache
- Page table
- Process descriptor
Queueing Discipline, queues packets before they are sent out to the network device, enforces QoS requirements, provides traffic shaping and prioritizing capabilities.
Read Copy Update, a mechanism for SMPSynchronisation
a lock mechanism that works per process context, see SMPSynchronisation
the part of the kernel that chooses a suitable process to run on the cpu, see the schedule() function.
- Shared/Paged Socket Buffer
- Slab cache
- Socket Buffer
(also: skb or sk_buff) data structure used to hold the data and attributes of a network packet. See http://www.linux-foundation.org/en/Net:SK_Buff and http://vger.kernel.org/~davem/skb.html for details.
- Spin lock
a simple SMP lock, see SMPSynchronisation
- Swap token
a token to temporarily protect a process from pageout, an alternative approach to memory scheduling, thrashing control. See the Token Based Thrashing Control paper by Song Jiang and the Linux-MM wiki.
- System call
A pair of instructions on Pentium2+ that replace older INT instruction based syscall mechanism. See http://articles.manugarg.com/systemcallinlinux2_6.html
the page replacement algorithm used by the Linux 2.6 kernel, based on the ideas behind the 2Q page replacement algorithm, also see the AdvancedPageReplacement page.
Virtual Dynamically-linked Shared Object, a kernel-provided shared library that helps userspace perform a few kernel actions without the overhead of a system call, as well as automatically choosing the most efficient syscall mechanism. Also called the "vsyscall page".
- Virtual memory
- Vsyscall page
A paravirtualisation engine for Linux, an efficient way to run multiple Linux OSes on one computer. Also runs BSD, Plan9 and other OSes. (See website for more information.)
Yellowdog Updater Modified yum is an interactive, rpm based, package manager. It can automatically perform system updates, including dependency analysis and absolete processing based on "repository" metadata. It can also perform installation of new packages, removal of old packages and perform queries on the installed and/or available packages among many other commands/services. Yum is similar to other high level package managers like apt-get and smart.